# 外部候选材料

[返回有排名的队列](./QUEUE.md) · [说明](./README.md)

以下记录尚未进入正式排名。序号仅便于浏览，保留原 catalog 的记录顺序，不表示优先级。

| 顺序 | 材料 | 档位 | 读取范围与备注 |
| --- | --- | --- | --- |
| 1 | [SWE-Bench Pro Public Dataset / Scale AI](<https://labs.scale.com/leaderboard/swe_bench_pro_public>) | A | 沿用既有阅读记录 |
| 2 | [SEVA: Self-Evolving Verification Agent with Process Reward for Fact Attribution](<https://arxiv.org/abs/2606.29713>) | A | 沿用既有阅读记录 |
| 3 | [LongCat-2.0](<https://github.com/meituan-longcat/LongCat-2.0>) · [来源 2](<https://huggingface.co/meituan-longcat/LongCat-2.0>) | A | 沿用既有阅读记录 |
| 4 | [AReaL 2.0 / Next-Generation Agentic RL Systems](<https://arxiv.org/pdf/2607.01120>) · [来源 2](<https://github.com/areal-project/AReaL>) | A | 沿用既有阅读记录 |
| 5 | [AutoMem: Automated Learning of Memory as a Cognitive Skill](<https://arxiv.org/abs/2607.01224>) | A | 沿用既有阅读记录 |
| 6 | [Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity](<https://www.microsoft.com/en-us/research/publication/memora-a-harmonic-memory-representation-balancing-abstraction-and-specificity/>) | A | 沿用既有阅读记录 |
| 7 | [NVIDIA ASPIRE: Agentic Skills Discovery for Robotics](<https://research.nvidia.com/labs/gear/aspire/>) | A | 沿用既有阅读记录 |
| 8 | [Safari MCP server for web developers](<https://webkit.org/blog/18136/introducing-the-safari-mcp-server-for-web-developers/>) | A | 沿用既有阅读记录 |
| 9 | [Manufact MCP Cloud / mcp-use production lifecycle](<https://manufact.com/>) | A | 沿用既有阅读记录 |
| 10 | [Cloudflare AI traffic options: Search / Agent / Training crawler split](<https://blog.cloudflare.com/content-independence-day-ai-options/>) | A | 沿用既有阅读记录 |
| 11 | [OpenSquilla 0.4.0 coding mode and verification workflow](<https://opensquilla.ai/news/0.4.0-release/>) · [来源 2](<https://github.com/opensquilla/opensquilla>) | A | 沿用既有阅读记录 |
| 12 | [Raft Blog：Agent-native workspace / AX / human-agent team product language](<https://raft.build/zh-cn/resources/blog/>) | A | 沿用既有阅读记录 |
| 13 | [AutoPass：Evidence-Guided LLM Agents for Compiler Performance Tuning](<https://arxiv.org/abs/2606.20373>) | A | 沿用既有阅读记录 |
| 14 | [PowerAgentBench-Dyn：A Benchmark for Agentic AI in Power System Dynamic Studies](<https://arxiv.org/abs/2606.20401>) | A | 沿用既有阅读记录 |
| 15 | [RetailBench：Benchmarking long horizon reasoning and coherent decision making of LLM agents in realistic retail environments](<https://arxiv.org/html/2606.15862v1>) | A | 沿用既有阅读记录 |
| 16 | [Multi-LCB：Extending LiveCodeBench to Multiple Programming Languages](<https://arxiv.org/abs/2606.20517>) | A | 沿用既有阅读记录 |
| 17 | [Statistical Foundations of LLM-based A/B Testing: A Surrogacy Framework for Human Causal Inference](<https://arxiv.org/abs/2606.17165>) | A | 沿用既有阅读记录 |
| 18 | [Apodex-1.0：verification-centric deep-research agent team / AgentOS / AgentHarness](<http://xhslink.com/o/1TKKmKWIu5x>) · [来源 2](<https://github.com/ApodexAI/AgentHarness>) · [来源 3](<https://huggingface.co/collections/apodex/apodex-1>) | A | 沿用既有阅读记录 |
| 19 | [zartbot：用 Agentic Workflow 做大模型全栈研究 / AI-infra auto-research](<https://mp.weixin.qq.com/s/5wm1tgR7d9LNk2ZIltX3CA>) | A | 沿用既有阅读记录 |
| 20 | [Oriol Vinyals / Gemini Co-Lead：World Models、Agent Scaffolding、Memory 与 Post-Training RL 路线访谈](<https://mp.weixin.qq.com/s/IjdO7OjxogK_W_aps6UJqA>) | A | 沿用既有阅读记录 |
| 21 | [MemDecoder：把 memory composition 做成 autoregressive index decoding](<http://xhslink.com/o/1mQp3aTSBjg>) | A | 沿用既有阅读记录 |
| 22 | [MemEvolve：把 agent memory architecture 本身纳入 meta-evolution](<http://xhslink.com/o/5HwWelHrLws>) · [来源 2](<https://arxiv.org/abs/2512.18746>) · [来源 3](<https://github.com/bingreeky/MemEvolve>) | A | 沿用既有阅读记录 |
| 23 | [SimpleMem：semantic lossless compression + adaptive retrieval 的 lifelong memory substrate](<http://xhslink.com/o/5HwWelHrLws>) · [来源 2](<https://arxiv.org/abs/2601.02553>) · [来源 3](<https://openreview.net/forum?id=CMveUVer0m>) · [来源 4](<https://github.com/aiming-lab/SimpleMem>) | A | 沿用既有阅读记录 |
| 24 | [MemOCR：把 memory 渲染成 layout-aware visual context 来分配信息密度](<http://xhslink.com/o/5HwWelHrLws>) · [来源 2](<https://arxiv.org/abs/2601.21468>) · [来源 3](<https://github.com/meituan/MemOCR>) | A | 沿用既有阅读记录 |
| 25 | [Darwinian Memory：GUI agent 的 utility-driven natural selection memory](<http://xhslink.com/o/5HwWelHrLws>) · [来源 2](<https://arxiv.org/abs/2601.22528>) | A | 沿用既有阅读记录 |
| 26 | [EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective](<https://arxiv.org/abs/2605.18421>) · [来源 2](<https://github.com/DSAIL-Memory/EvoMemBench>) | A | 沿用既有阅读记录 |
| 27 | [MemMark: State-Evolution Attribution Watermarking for Agent Long-Term Memory Systems](<https://arxiv.org/abs/2605.25002>) | A | 沿用既有阅读记录 |
| 28 | [Portable Agent Memory: A Protocol for Cryptographically-Verified Memory Transfer Across Heterogeneous AI Agents](<https://arxiv.org/abs/2605.11032>) | A | 沿用既有阅读记录 |
| 29 | [SemiAnalysis / 腾讯科技：AI Dark Output 与 GDP 统计黑洞](<https://mp.weixin.qq.com/s/ChYdriATPMeqGfjSaK7MCw>) | A | 沿用既有阅读记录 |
| 30 | [ACE: Agentic Context Engineering for Self-Improving Language Models](<https://openreview.net/forum?id=eC4ygDs02R>) | A | 沿用既有阅读记录 |
| 31 | [ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory](<https://openreview.net/forum?id=jL7fwchScm>) | A | 沿用既有阅读记录 |
| 32 | [XENON: Experience-based Knowledge Correction for Robust Planning in Minecraft](<https://openreview.net/forum?id=N22lDHYrXe>) | A | 沿用既有阅读记录 |
| 33 | [FlowSearcher: Synthesizing Memory-Guided Agentic Workflows for Web Information Seeking](<https://openreview.net/forum?id=34v7DVz2l0>) | A | 沿用既有阅读记录 |
| 34 | [AgentFlow: In-the-Flow Agentic System Optimization for Effective Planning and Tool Use](<https://openreview.net/forum?id=Mf5AleTUVK>) | A | 沿用既有阅读记录 |
| 35 | [MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent](<https://openreview.net/forum?id=k5nIOvYGCL>) | A | 沿用既有阅读记录 |
| 36 | [REMem: Reasoning with Episodic Memory in Language Agent](<https://openreview.net/forum?id=fugnQxbvMm>) | A | 沿用既有阅读记录 |
| 37 | [ReMemR1: Look Back to Reason Forward: Revisitable Memory for Long-Context LLM Agents](<https://openreview.net/forum?id=1cymflI2Lh>) | A | 沿用既有阅读记录 |
| 38 | [GraphPlanner: Graph Memory-Augmented Agentic Routing for Multi-Agent LLMs](<https://openreview.net/forum?id=ZdGB7MNQDT>) | A | 沿用既有阅读记录 |
| 39 | [GLoW: Dual-Scale World Memory for LLM Agents towards Hard-Exploration Problems](<https://openreview.net/forum?id=bH5uHIVtTe>) | A | 沿用既有阅读记录 |
| 40 | [Improving Code Localization with Repository Memory](<https://openreview.net/forum?id=8yjWLJy2eX>) | A | 沿用既有阅读记录 |
| 41 | [TokMem: One-Token Procedural Memory for Large Language Models](<https://openreview.net/forum?id=RWjEf9PdiJ>) | A | 沿用既有阅读记录 |
| 42 | [Sculptor: Empowering LLMs with Cognitive Agency via Active Context Management](<https://openreview.net/forum?id=HPeiH7da0Z>) | A | 沿用既有阅读记录 |
| 43 | [DataMind: Scaling Generalist Data-Analytic Agents](<https://openreview.net/forum?id=5PxFqpIYWC>) | A | 沿用既有阅读记录 |
| 44 | [MemGen: Weaving Generative Latent Memory for Self-Evolving Agents](<https://openreview.net/forum?id=vI56m4Iu4e>) | A | 沿用既有阅读记录 |
| 45 | [Agent-World：可扩展真实环境合成与自进化 Agent 训练](<https://arxiv.org/abs/2604.18292>) | A | 公开摘要页已核验；本次核验公开论文页面；技术细节沿用既有阅读记录。 |
| 46 | [LangMem / LangGraph long-term memory docs](<https://github.com/langchain-ai/langmem>) | B | 沿用既有阅读记录 |
| 47 | [Microsoft GraphRAG / graph-based retrieval baseline](<https://microsoft.github.io/graphrag/>) | B | 沿用既有阅读记录 |
| 48 | [Hands-On Modern RL](<https://walkinglabs.github.io/hands-on-modern-rl/>) | A | 沿用既有阅读记录 |
| 49 | [AgentLongBench：用 environment rollout 测 long-context agent，而不是静态 retrieval](<https://arxiv.org/abs/2601.20730>) · [来源 2](<https://github.com/euReKa025/AgentLongBench>) · [来源 3](<https://huggingface.co/papers/2601.20730>) | A | 沿用既有阅读记录 |
| 50 | [PlugMem：把 raw experience 压成 knowledge-centric memory graph](<http://xhslink.com/o/1mQp3aTSBjg>) · [来源 2](<https://arxiv.org/abs/2603.03296>) · [来源 3](<https://github.com/TIMAN-group/PlugMem>) | A | 沿用既有阅读记录 |
| 51 | [SGLang Disaggregated Serving：P/D split 与 KV transfer 的另一套工程参照](<https://docs.nvidia.com/dynamo/archive/0.7.0/backends/sglang/sglang-disaggregation.html>) | B | 沿用既有阅读记录 |
| 52 | [Continuum：multi-turn agent scheduling with KV cache TTL](<https://arxiv.org/abs/2511.02230>) · [来源 2](<https://openreview.net/forum?id=sVzK0LC9pn>) | A | 沿用既有阅读记录 |
| 53 | [Agentic KV-cache radar：PrefillShare / TokenDance / CONCUR](<https://arxiv.org/abs/2602.12029>) · [来源 2](<https://arxiv.org/abs/2604.03143>) · [来源 3](<https://arxiv.org/abs/2601.22705>) | B | 沿用既有阅读记录 |
| 54 | [ThunderAgent：program-aware agentic inference system](<https://arxiv.org/abs/2602.13692>) · [来源 2](<https://github.com/ThunderAgent-org/ThunderAgent>) | A | 沿用既有阅读记录 |
| 55 | [Cloudflare Code Mode MCP：把工具面压成可执行代码接口](<https://blog.cloudflare.com/code-mode-mcp/>) · [来源 2](<https://github.com/cloudflare/mcp>) | A | 沿用既有阅读记录 |
| 56 | [OpenAI Responses API WebSocket：agent loop state reuse as API/runtime optimization](<https://openai.com/index/speeding-up-agentic-workflows-with-websockets/>) | A | 沿用既有阅读记录 |
| 57 | [Don't Break the Cache：long-horizon agent prompt caching 的策略边界](<https://arxiv.org/abs/2601.06007>) | A | 沿用既有阅读记录 |
| 58 | [Datadog State of AI Engineering：生产 agent 的 prompt/cache/observability 统计锚点](<https://www.datadoghq.com/state-of-ai-engineering/>) | A | 沿用既有阅读记录 |
| 59 | [NVIDIA Extreme Co-Design for Agentic Systems：agent token economics / sub-agent / compaction 公开案例](<https://developer.nvidia.com/blog/building-for-the-rising-complexity-of-agentic-systems-with-extreme-co-design/>) | A | 沿用既有阅读记录 |
| 60 | [LangSmith / Langfuse observability model：trace / run / session / feedback / observation](<https://docs.langchain.com/langsmith/observability-concepts>) | A | 沿用既有阅读记录 |
| 61 | [AgentTrace / AgentSight：agent observability beyond LLM tracing](<https://arxiv.org/abs/2602.10133>) · [来源 2](<https://arxiv.org/abs/2603.14688>) · [来源 3](<https://arxiv.org/abs/2508.02736>) | B | 沿用既有阅读记录 |
| 62 | [Lilian Weng: Why We Think](<https://lilianweng.github.io/posts/2025-05-01-thinking/>) | A | 沿用既有阅读记录 |
| 63 | [ezyang: OSS code review, in the era of LLMs](<https://blog.ezyang.com/2026/04/oss-code-review-in-the-era-of-llms/>) | A | 沿用既有阅读记录 |
| 64 | [Workflow-aware serving radar：Helium / ForkKV / ToolCacheAgent](<https://arxiv.org/abs/2603.16104>) · [来源 2](<https://arxiv.org/abs/2604.06370>) · [来源 3](<https://openreview.net/forum?id=tX3YcbNa5w>) | B | 沿用既有阅读记录 |
| 65 | [GPU MODE Resource Stream / Reference Kernels](<https://github.com/gpu-mode/resource-stream>) · [来源 2](<https://github.com/gpu-mode/reference-kernels>) · [来源 3](<https://github.com/gpu-mode>) | B | 沿用既有阅读记录 |
| 66 | [PyTorch Dev Discuss: performance / deployment categories](<https://dev-discuss.pytorch.org/>) | B | 沿用既有阅读记录 |
| 67 | [Claw-Eval-Live：用 live workflow demand signal 校准 agent benchmark](<https://claw-eval-live.github.io/>) · [来源 2](<https://huggingface.co/papers/2604.28139>) | A | 沿用既有阅读记录 |
| 68 | [Claw-Eval：Pass@k / Pass^k 与 trajectory-aware grading 的 agent eval 锚点](<https://arxiv.org/abs/2604.06132>) · [来源 2](<https://github.com/claw-eval/claw-eval>) | A | 沿用既有阅读记录 |
| 69 | [AgentSwing：parallel branch + lookahead routing 的 context management 机制线索](<https://arxiv.org/abs/2603.27490>) | A | 沿用既有阅读记录 |
| 70 | [SkillFlow：lifelong skill discovery / repair / library evolution](<https://arxiv.org/abs/2604.17308>) | A | 沿用既有阅读记录 |
| 71 | [Live-Evo：online self-evolving memory from continuous feedback](<https://arxiv.org/abs/2602.02369>) | A | 沿用既有阅读记录 |
| 72 | [MemSkill：learnable and evolvable memory skills](<https://arxiv.org/abs/2602.02474>) | A | 沿用既有阅读记录 |
| 73 | [MemoryCD：cross-domain user memory benchmark from real behavior](<https://arxiv.org/abs/2603.25973>) | A | 沿用既有阅读记录 |
| 74 | [A-Evolve：agent evolution as infrastructure](<https://arxiv.org/abs/2602.00359>) · [来源 2](<https://github.com/A-EVO-Lab/a-evolve>) | B | 沿用既有阅读记录 |
| 75 | [agent-eval：Inspect-format eval suite and leaderboard tooling](<https://pypi.org/project/agent-eval/>) | B | 沿用既有阅读记录 |
| 76 | [Anthropic: Harness design for long-running application development](<https://www.anthropic.com/engineering/harness-design-long-running-apps>) | S | 沿用既有阅读记录 |
| 77 | [10 篇论文拆解 Skill + 自进化的技术路线](<https://mp.weixin.qq.com/s/k2vjcm5ctSdCYXRDko9CZQ>) | S | 沿用既有阅读记录 |
| 78 | [火山养“龙虾”日志：Self-Improving Skill 让 AI 学会“自我进化”](<https://mp.weixin.qq.com/s/qcX68XwJliiSmxzKXKoscA>) | S | 沿用既有阅读记录 |
| 79 | [Context Engineering 2.0: The Context of Context Engineering](<https://arxiv.org/abs/2510.26493>) · [来源 2](<https://arxiv.org/pdf/2510.26493>) | A | 沿用既有阅读记录 |
| 80 | [Anthropic: Effective context engineering for AI agents](<https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents>) | A | 沿用既有阅读记录 |
| 81 | [Prompting Guide: Context Engineering Guide](<https://www.promptingguide.ai/guides/context-engineering-guide.en>) | B | 沿用既有阅读记录 |
| 82 | [微信导读：超越 Prompt 和 RAG，「上下文工程」成了 Agent 核心胜负手](<https://mp.weixin.qq.com/s/3t4PjpZcMVU1wCO0ThUs2A>) | B | 沿用既有阅读记录 |
| 83 | [Memex(RL)：stable index + full-fidelity evidence 的 indexed experience memory](<https://arxiv.org/abs/2603.04257>) | S | 沿用既有阅读记录 |
| 84 | [AgentRR：Record & Replay 范式、multi-level experience 与 check function](<https://arxiv.org/abs/2505.17716>) | A | 沿用既有阅读记录 |
| 85 | [A-MEM：Agentic Memory 的动态索引、链接与 memory evolution](<https://arxiv.org/abs/2502.12110>) · [来源 2](<https://github.com/WujiangXu/A-mem>) | S | 沿用既有阅读记录 |
| 86 | [Mem0：生产级 agent memory 的 hybrid search、entity linking 与 token/latency 约束](<https://arxiv.org/abs/2504.19413>) · [来源 2](<https://github.com/mem0ai/mem0>) | A | 沿用既有阅读记录 |
| 87 | [ProRAG：RAG 中 process reward 与 step-level credit assignment](<https://arxiv.org/abs/2601.21912>) | A | 沿用既有阅读记录 |
| 88 | [AgeMem：把长期/短期记忆管理整合进 agent policy](<https://arxiv.org/abs/2601.01885>) | A | 沿用既有阅读记录 |
| 89 | [Mem-T：Memory Operation Tree 与 MoT-GRPO 的 reward densification](<https://arxiv.org/abs/2601.23014>) | A | 沿用既有阅读记录 |
| 90 | [UMA / Learning to Remember：end-to-end memory agent、CRUD 与 Ledger-QA](<https://arxiv.org/abs/2602.18493>) | A | 沿用既有阅读记录 |
| 91 | [Memory-R1：用 RL 学习 ADD/UPDATE/DELETE/NOOP 与 memory utilization](<https://arxiv.org/abs/2508.19828>) | A | 沿用既有阅读记录 |
| 92 | [OCR-Memory：用视觉锚点保存长程 agent 轨迹并按需转录原文](<https://arxiv.org/abs/2604.26622>) | A | 沿用既有阅读记录 |
| 93 | [MemOS：把 memory 当作可管理系统资源的 Memory OS](<https://arxiv.org/abs/2507.03724>) · [来源 2](<https://github.com/MemTensor/MemOS>) | A | 沿用既有阅读记录 |
| 94 | [Engram：DeepSeek 条件记忆，把静态查表从动态推理中拆出来](<https://arxiv.org/abs/2601.07372>) · [来源 2](<https://github.com/deepseek-ai/Engram>) | B | 沿用既有阅读记录 |
| 95 | [LLMs-augmented Contextual Bandit：LLM encoder + bandit 的基础概念锚点](<https://arxiv.org/abs/2311.02268>) | B | 沿用既有阅读记录 |
| 96 | [MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks](<http://xhslink.com/o/5HwWelHrLws>) · [来源 2](<https://arxiv.org/abs/2602.16313>) | S | 沿用既有阅读记录 |
| 97 | [AgentHER: Hindsight Experience Replay for LLM Agent Trajectory Relabeling](<https://arxiv.org/abs/2603.21357>) | S | 沿用既有阅读记录 |
| 98 | [ProRL Agent: Rollout-as-a-Service for RL Training of Multi-Turn LLM Agents](<https://arxiv.org/abs/2603.18815>) | S | 沿用既有阅读记录 |
| 99 | [SeeUPO: Sequence-Level Agentic-RL with Convergence Guarantees](<https://arxiv.org/abs/2602.06554>) | S | 沿用既有阅读记录 |
| 100 | [Stabilizing Reinforcement Learning with LLMs: Formulation and Practices](<https://arxiv.org/abs/2512.01374>) | S | 沿用既有阅读记录 |
| 101 | [Welcome to the Era of Experience：Sutton / Silver 的经验时代与 Agentic RL 主张](<https://mp.weixin.qq.com/s/Rl-YUOIMxmpw_Ca6vf2YxA>) | A | 沿用既有阅读记录 |
| 102 | [DeepResearcher：真实 Web 环境中的 deep research agent RL](<https://arxiv.org/abs/2504.03160>) · [来源 2](<https://github.com/GAIR-NLP/DeepResearcher>) | A | 公开摘要页已核验；本次核验公开论文页面；技术细节沿用既有阅读记录。 |
| 103 | [Rethinking Memory in LLM-based Agents：AI 记忆表示、六大原子操作与 Memory Compass](<https://mp.weixin.qq.com/s/WeOf-Eqc_C0m94aXhpeECg>) · [来源 2](<https://arxiv.org/abs/2505.00675>) · [来源 3](<https://github.com/Elvin-Yiming-Du/Survey_Memory_in_AI>) | A | 沿用既有阅读记录 |
| 104 | [WebXSkill + Browserbase Skills：可执行 Skill 作为 Web Agent substrate](<https://arxiv.org/abs/2604.13318>) · [来源 2](<https://github.com/browserbase/skills>) | S | 沿用既有阅读记录 |
| 105 | [τ²-Bench: Evaluating Conversational Agents in a Dual-Control Environment](<https://arxiv.org/abs/2506.07982>) | A | 沿用既有阅读记录 |
| 106 | [agentevals：基于 OpenTelemetry trace 的本地离线 Agent 评测](<https://github.com/agentevals-dev/agentevals>) | A | 沿用既有阅读记录 |
| 107 | [BudgetMem: Learning Query-Aware Budget-Tier Routing for Runtime Agent Memory](<https://arxiv.org/abs/2602.06025>) | A | 沿用既有阅读记录 |
| 108 | [MemRouter: Memory-as-Embedding Routing for Long-Term Conversational Agents](<https://papers.cool/arxiv/2605.00356>) | A | 沿用既有阅读记录 |
| 109 | [ParetoBandit: Budget-Paced Adaptive Routing for Non-Stationary LLM Serving](<https://arxiv.org/abs/2604.00136>) | A | 沿用既有阅读记录 |
| 110 | [AgentTrace: A Structured Logging Framework for Agent System Observability](<https://arxiv.org/abs/2602.10133>) | A | 沿用既有阅读记录 |
| 111 | [o11y-bench：AI Agent 的 Observability 任务 benchmark](<https://o11ybench.ai/>) | A | 沿用既有阅读记录 |
| 112 | [ATBench：trajectory-level agent safety benchmark](<https://arxiv.org/abs/2604.02022>) | A | 沿用既有阅读记录 |
| 113 | [mem9：OpenClaw 云端永续记忆、Memory Space 与 ContextEngine 生命周期接口](<https://mp.weixin.qq.com/s/EASFARQBcyzFOLXEP4wE6A>) · [来源 2](<https://github.com/mem9-ai/mem9>) | A | 沿用既有阅读记录 |
| 114 | [ArkClaw 十大热门 Skills：技能生态、加载优先级、多 Agent 绑定与 APM 观测](<https://mp.weixin.qq.com/s/RbAYzE224G1nrHSqxyjGqg>) | A | 沿用既有阅读记录 |
| 115 | [How To Be A World-Class Agentic Engineer：保持简单、上下文管理与 contract-driven session](<https://x.com/systematicls/status/2028814227004395561>) | A | 沿用既有阅读记录 |
| 116 | [Agent Skills Prompt Injection：Skill 文件带来的现实攻击面](<https://arxiv.org/abs/2510.26328>) | A | 沿用既有阅读记录 |
| 117 | [Claw Code / Claude Code 开源 Rust agent harness 解读](<https://zread.ai/instructkr/claude-code/1-overview>) | A | 沿用既有阅读记录 |
| 118 | [Multi-Agent 是伪技术路线？OpenClaw subagents 与多 Agent 适用边界](<http://xhslink.com/o/15v8SNPxd9q>) · [来源 2](<https://arxiv.org/abs/2512.08296>) | A | 沿用既有阅读记录 |
| 119 | [Why Do Multi-Agent LLM Systems Fail?：MAST 多智能体失败分类与 trace 数据集](<https://arxiv.org/abs/2503.13657>) · [来源 2](<https://github.com/multi-agent-systems-failure-taxonomy/MAST>) | A | 沿用既有阅读记录 |
| 120 | [Claude Code Superpowers：结构化软件工程技能框架](<https://claude.com/plugins/superpowers>) · [来源 2](<https://github.com/obra/superpowers>) | A | 沿用既有阅读记录 |
| 121 | [EvoMap / Evolver：GEP 能力进化资产与跨 Agent 经验共享](<https://github.com/EvoMap/evolver>) | A | 沿用既有阅读记录 |
| 122 | [Agent Harness Engineering: A Survey](<http://xhslink.com/o/2yswRF03SIa>) · [来源 2](<https://openreview.net/pdf?id=eONq7FdiHa>) · [来源 3](<https://github.com/Picrew/awesome-agent-harness>) | A | 沿用既有阅读记录；旧记录含已读备注但受管生命周期仍为 candidate；保留状态冲突，待单独核对。 |
| 123 | [Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents](<https://arxiv.org/abs/2604.10674>) | A | 沿用既有阅读记录 |
| 124 | [SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization](<https://arxiv.org/abs/2604.02268>) · [来源 2](<https://github.com/ZJU-REAL/SkillZero>) | A | 沿用既有阅读记录 |
| 125 | [SkillOS: Learning Skill Curation for Self-Evolving Agents](<http://xhslink.com/o/1wkuGrStMtK>) · [来源 2](<https://arxiv.org/abs/2605.06614>) | A | 沿用既有阅读记录 |
| 126 | [Voyager: An Open-Ended Embodied Agent with Large Language Models](<https://arxiv.org/abs/2305.16291>) · [来源 2](<https://github.com/MineDojo/Voyager>) | A | 沿用既有阅读记录 |
| 127 | [Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers](<https://arxiv.org/abs/2603.07670>) | A | 沿用既有阅读记录 |
| 128 | [AMV-L: Lifecycle-Managed Agent Memory for Tail-Latency Control in Long-Running LLM Systems](<https://arxiv.org/abs/2603.04443>) | A | 沿用既有阅读记录 |
| 129 | [Coding Agents are Effective Long-Context Processors](<https://arxiv.org/abs/2603.20432>) | A | 沿用既有阅读记录 |
| 130 | [On Data Engineering for Scaling LLM Terminal Capabilities：Terminal Agent 数据工程](<https://www.bilibili.com/video/BV1zRP7zVEz2>) · [来源 2](<https://arxiv.org/abs/2602.21193>) | A | 沿用既有阅读记录 |
| 131 | [A Guide to Gen AI / LLM Vibecoding for Expert Programmers](<https://www.stochasticlifestyle.com/a-guide-to-gen-ai-llm-vibecoding-for-expert-programmers/>) | A | 沿用既有阅读记录 |
| 132 | [Exgentic Open Agent Leaderboard：跨 benchmark 的统一 Agent 评测协议](<https://www.exgentic.ai/>) | A | 沿用既有阅读记录 |
| 133 | [AgencyBench：1M-token long-horizon autonomous agent benchmark](<https://github.com/GAIR-NLP/AgencyBench>) | A | 沿用既有阅读记录 |
| 134 | [Agent Replay：local-first desktop evals / observability / memory](<https://github.com/agentreplay/agentreplay>) | B | 沿用既有阅读记录 |
| 135 | [OpenAI: Speeding up agentic workflows with WebSockets in the Responses API](<https://openai.com/index/speeding-up-agentic-workflows-with-websockets/>) | A | 沿用既有阅读记录 |
| 136 | [Glean Waldo：专门的 agentic search model 做检索规划](<https://www.glean.com/jp/blog/waldo-launch>) | A | 沿用既有阅读记录 |
| 137 | [CC Pocket](<https://github.com/K9i-0/ccpocket/blob/main/README.zh-CN.md>) | A | 沿用既有阅读记录 |
| 138 | [ArkClaw 动手实验上新：财经、评论洞察、财报、视频剪辑、播客生成](<https://mp.weixin.qq.com/s/B9GM4CfMOuKv4eUokdwGRg>) | A | 沿用既有阅读记录 |
| 139 | [火山联网搜索 Skill：ArkClaw / OpenClaw 的 Agent 原生搜索能力](<https://mp.weixin.qq.com/s/HN_50ugvFQtVujRkrbJHGQ>) | A | 沿用既有阅读记录 |
| 140 | [被 OpenClaw 推上风口的飞书：企业 Agent 落地、工作空间容器与组织协作变迁](<https://mp.weixin.qq.com/s/2eU_5442rPe2YKWiXDR_5Q>) | A | 沿用既有阅读记录 |
| 141 | [碎片化编程与手机访问 Claude Code / OpenClaw](<https://mp.weixin.qq.com/s/Oy_ogkPBJSNbkEi8ED3FLA>) | A | 沿用既有阅读记录 |
| 142 | [Agent-Reach：开源本地互联网工具脚手架](<https://github.com/Panniantong/Agent-Reach>) | A | 沿用既有阅读记录 |
| 143 | [Amazon SynerGen：搜推召排一体化的生成式推荐框架](<https://mp.weixin.qq.com/s/3B0s7Gv73NPZTLjEOn_ZNA>) · [来源 2](<https://arxiv.org/abs/2509.21777>) | A | 沿用既有阅读记录 |
| 144 | [快手 TagCF：用户角色、行为逻辑图与推荐系统显式破茧](<https://mp.weixin.qq.com/s/9cAw5GWYRW8vZs7S2io7KA>) · [来源 2](<https://arxiv.org/abs/2505.10940>) · [来源 3](<https://github.com/Code2Q/TagCF>) | A | 沿用既有阅读记录 |
| 145 | [DualGR：快手长短期兴趣建模的生成式召回实践](<https://mp.weixin.qq.com/s/fh7sW6yYBlNldoWh9v7GIg>) · [来源 2](<https://arxiv.org/abs/2511.12518>) | A | 沿用既有阅读记录 |
| 146 | [LEMUR：大规模端到端多模态推荐](<https://arxiv.org/abs/2511.10962>) | A | 沿用既有阅读记录 |
| 147 | [Don’t Waste It：用结构化 Human Priors 指导生成式推荐多头解码](<https://arxiv.org/html/2511.10492v3>) · [来源 2](<https://github.com/zhykoties/Multi-Head-Recommendation-with-Human-Priors>) | A | 沿用既有阅读记录 |
| 148 | [SSCTL：多领域推荐中的数据不均衡与半监督迁移](<https://arxiv.org/pdf/2511.01404>) | A | 沿用既有阅读记录 |
| 149 | [GNOLR：多隐式反馈的有序偏好统一嵌入与推荐召回简化](<https://zhuanlan.zhihu.com/p/1913992712039957453>) · [来源 2](<https://arxiv.org/abs/2505.20900>) · [来源 3](<https://github.com/FuCongResearchSquad/GNOLR>) | A | 沿用既有阅读记录 |
| 150 | [搜索推荐推理合一 / LLM 混合 ID 生成：NEO at Spotify](<http://xhslink.com/o/7RUl8KFUtko>) · [来源 2](<https://arxiv.org/abs/2603.17533>) | A | 沿用既有阅读记录 |
| 151 | [Google STATIC：向量化 Trie，加速生成式召回约束解码](<https://mp.weixin.qq.com/s/La6KMPgKPZ5gWVLh1_0gbg>) · [来源 2](<https://arxiv.org/abs/2602.22647>) · [来源 3](<https://github.com/youtube/static-constraint-decoding>) | A | 沿用既有阅读记录 |
| 152 | [CollectiveKV：序列推荐中的跨用户 KV cache 共享与压缩](<https://mp.weixin.qq.com/s/80q5go7AVMUxZG6JPWQjRw>) · [来源 2](<https://arxiv.org/abs/2601.19178>) | A | 沿用既有阅读记录 |
| 153 | [Hiformer：Google Play 推荐排序中的异构特征交互 Transformer](<https://arxiv.org/abs/2311.05884>) · [来源 2](<https://huggingface.co/papers/2311.05884>) | A | 沿用既有阅读记录 |
| 154 | [HyFormer: Revisiting the Roles of Sequence Modeling and Feature Interaction in CTR Prediction](<https://arxiv.org/abs/2601.12681>) | A | 沿用既有阅读记录 |
| 155 | [X / Twitter For You 推荐算法：Grok-based Transformer、两塔召回与 Candidate Isolation](<https://mp.weixin.qq.com/s/3tvsymPAHEPUzAT1IXj9dQ>) · [来源 2](<https://github.com/xai-org/x-algorithm>) | A | 沿用既有阅读记录 |
| 156 | [Meta SilverTorch：召回与粗排全链路 GPU 算子化](<https://mp.weixin.qq.com/s/yvaTutx8PiMoS184ts_Tew>) · [来源 2](<https://arxiv.org/abs/2511.14881>) | A | 沿用既有阅读记录 |
| 157 | [LORE：阿里广告搜索相关性大模型与垂直 RL 经验](<https://mp.weixin.qq.com/s/wly9f8dWTkIp7pYM8j-ANg>) · [来源 2](<https://arxiv.org/abs/2512.03025>) | A | 沿用既有阅读记录 |
| 158 | [谈谈 RL Infra / FlashRL：训练、推理、调度、权重同步与 observability](<http://xhslink.com/o/7IQP0wPYMcW>) · [来源 2](<https://github.com/lastweek/FlashRL>) | A | 沿用既有阅读记录 |
| 159 | [Ilya Sutskever：An Observation on Generalization / 用压缩视角理解无监督学习](<http://xhslink.com/o/2IdP9c96xLI>) | A | 沿用既有阅读记录 |
| 160 | [Muon 优化器：矩阵正交化更新、LLM 训练可扩展性与 Kimi K2 的 MuonClip](<https://github.com/KellerJordan/Muon>) · [来源 2](<https://arxiv.org/abs/2502.16982>) · [来源 3](<https://arxiv.org/abs/2507.20534>) | A | 沿用既有阅读记录 |
| 161 | [DeepSeek-V4 Technical Report：Million-Token Context、CSA/HCA、mHC、Muon 与长程 Agent 能力](<https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/DeepSeek_V4.pdf>) · [来源 2](<https://huggingface.co/collections/deepseek-ai/deepseek-v4>) | A | 沿用既有阅读记录 |
| 162 | [PyGraph：让 PyTorch 的 CUDA Graph 优化更高效](<https://mp.weixin.qq.com/s/DRnDL_dpyR793kEkzLgoWw>) · [来源 2](<https://arxiv.org/abs/2503.19779>) | A | 沿用既有阅读记录 |
| 163 | [PyTorch FlexAttention：用 score_mod / mask_mod 表达 Attention 变体并编译成高性能 Triton kernel](<https://pytorch.org/blog/flexattention/>) · [来源 2](<https://github.com/pytorch-labs/attention-gym>) | A | 沿用既有阅读记录 |
| 164 | [Intel / 龙蜥：xFasterTransformer、至强 AMX/HBM/CXL 与 CPU LLM 推理背景](<https://developer.aliyun.com/article/1462230>) | A | 沿用既有阅读记录 |
| 165 | [Look Ma, No Bubbles!：Llama-1B 低延迟 Megakernel 与 memory pipeline bubbles](<https://hazyresearch.stanford.edu/blog/2025-05-27-no-bubbles>) · [来源 2](<https://github.com/HazyResearch/Megakernels>) | A | 沿用既有阅读记录 |
| 166 | [CUDA Agent：基于真实性能反馈的 agentic RL CUDA 内核生成](<https://mp.weixin.qq.com/s/f9achGuLzEAID_NSviHDng>) · [来源 2](<https://arxiv.org/abs/2602.24286>) · [来源 3](<https://github.com/BytedTsinghua-SIA/CUDA-Agent>) | A | 沿用既有阅读记录 |
| 167 | [AutoResearch 自研高性能 GPU 算子 Flash Attention：MFU 42%、Opus 4.6、8 小时 25 轮迭代](<https://mp.weixin.qq.com/s/QwfzzGtDU7_B_mopjndMaw>) | A | 沿用既有阅读记录 |
| 168 | [DualPath：Agentic LLM Inference 的 KV-Cache 存储带宽瓶颈](<https://arxiv.org/pdf/2602.21548>) | A | 沿用既有阅读记录 |
| 169 | [MetaShuffling：Llama 4 MoE 推理的 shuffle-based Fused MoE kernel 与 Padding 避免](<https://mp.weixin.qq.com/s/MdztXkwIzw0ERTOVoCUz3g>) | A | 沿用既有阅读记录 |
| 170 | [Agent^2 RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training?](<https://arxiv.org/abs/2604.10547>) | A | 沿用既有阅读记录 |
| 171 | [少用 sense 挑战 math：如何把 post train 做好](<https://zhuanlan.zhihu.com/p/1995265459285694156>) | A | 沿用既有阅读记录 |
| 172 | [TTT-E2E：End-to-End Test-Time Training for Long Context](<https://mp.weixin.qq.com/s/OHJy0EEkYsnR_41OEHj5bA>) · [来源 2](<https://arxiv.org/abs/2512.23675>) | A | 沿用既有阅读记录 |
| 173 | [Nested Learning / Hope：多频率持续学习、优化器即联想记忆](<https://mp.weixin.qq.com/s/gse1I315nY2DtzG7qqaY6A>) · [来源 2](<https://arxiv.org/abs/2512.24695>) | A | 沿用既有阅读记录 |
| 174 | [林俊旸详解 Qwen 模型设计中的每一次取舍](<https://b23.tv/fObeCUa>) · [来源 2](<https://arxiv.org/abs/2309.16609>) | A | 沿用既有阅读记录 |
| 175 | [mHC: Manifold-Constrained Hyper-Connections](<https://arxiv.org/abs/2512.24880>) | A | 公开摘要页已核验；本次核验公开论文页面；技术细节沿用既有阅读记录。 |
| 176 | [MiniMax VTP：Towards Scalable Pre-training of Visual Tokenizers for Generation](<http://xhslink.com/o/1PR9ZDLnPPU>) · [来源 2](<https://arxiv.org/abs/2512.13687>) · [来源 3](<https://github.com/MiniMax-AI/VTP>) | A | 沿用既有阅读记录 |
| 177 | [Gated Attention for Large Language Models：Non-linearity, Sparsity, and Attention-Sink-Free](<https://arxiv.org/abs/2505.06708>) | A | 公开摘要页已核验；本次核验公开论文页面；技术细节沿用既有阅读记录。 |
| 178 | [Over-Tokenized Transformer：扩大输入词表作为稀疏 scaling 维度](<https://arxiv.org/pdf/2501.16975>) | A | 沿用既有阅读记录 |
| 179 | [MiniMax M2：为什么最终选择 full attention](<https://huggingface.co/blog/MiniMax-AI/why-did-m2-end-up-as-a-full-attention-model>) | A | 沿用既有阅读记录 |
| 180 | [王云鹤：Harness 是复杂优化问题，Agent = Models + Harness](<https://zhuanlan.zhihu.com/p/2038669387150927679>) | A | 沿用既有阅读记录 |
| 181 | [田渊栋访谈：Meta 裁员后，对 AI 研究、RL 与职业选择的思考](<https://www.douyin.com/video/7570984486024875306>) | A | 沿用既有阅读记录 |
| 182 | [WhynotTV Podcast #4：翁家翌谈 OpenAI、post-training、RL infra、工程能力与 impact](<https://b23.tv/jVcQ9ht>) | S | 沿用既有阅读记录 |
| 183 | [田渊栋：告别 OpenAI 小作文（二）未来会是什么样子](<https://zhuanlan.zhihu.com/p/1991073922217709984>) | S | 沿用既有阅读记录 |
| 184 | [Manus季逸超访谈解读（1）：Benchmark是所有AI公司的唯一护城河](<https://mp.weixin.qq.com/s/QW2Yljxrzm25yFOpln-DWA>) | S | 沿用既有阅读记录 |
| 185 | [小红书：从 L3 到 L8 分别需要怎么 vibe](<http://xhslink.com/o/8t2lrMuALTE>) | A | 沿用既有阅读记录 |
| 186 | [小红书：张咋啦 follow-builders skill / 关注 builders 而不是 KOL](<http://xhslink.com/o/9nzbS0yYGVa>) | A | 沿用既有阅读记录 |
| 187 | [Aha：AI 产品真正杠杆在后端，而不是对话框](<http://xhslink.com/o/4bIBtId6Xg2>) | A | 沿用既有阅读记录 |
| 188 | [如何做出好产品：张一鸣产品定律、推荐、AB 测试与务实文化](<https://mp.weixin.qq.com/s/s0hk-0nwtkdk-Dq1eeedAw>) | A | 沿用既有阅读记录 |
| 189 | [对话腾讯 ima 产品团队：有价值的产品，不需要告诉用户「这是智能体」](<https://zhuanlan.zhihu.com/p/1908164667077002671>) | A | 沿用既有阅读记录 |
| 190 | [锦秋基金臧天宇：我们看 AI 应用的思考脉络](<http://xhslink.com/o/8q8Wb4BjK8J>) | A | 沿用既有阅读记录 |
| 191 | [Agent 经济学：一人公司、交易成本、Evals 与管理成本转移](<https://mp.weixin.qq.com/s/mAPynqwsNKlQ07pM9lmv1g>) | B | 沿用既有阅读记录 |
| 192 | [Anthropic CEO Dario Amodei 访谈：Scaling Laws、动态数据、应用层机会与 AI 时代职业判断](<https://mp.weixin.qq.com/s/4dPYQ7eMIAj4ujc9BukgXQ>) | B | 沿用既有阅读记录 |
| 193 | [DeepMind AlphaGo 十年复盘：游戏化、搜索、验证器与科学发现](<https://mp.weixin.qq.com/s/HmDSO7M0bHVybQSAZQEcRg>) | B | 沿用既有阅读记录 |
| 194 | [Werner Vogels final keynote：Renaissance Developer 与 AI 时代工程师价值](<http://xhslink.com/o/2gZB0wmancp>) | B | 沿用既有阅读记录 |
| 195 | [SaaS + Agent 十人谈：盖雅工场的垂直 Agent 与结果付费判断](<https://mp.weixin.qq.com/s/68rBcE_veNsTMcIwlWdOrQ>) | B | 沿用既有阅读记录 |
| 196 | [AI Agent 产品落地五个坑：记忆、工具幻觉、调度、成本与执行精度](<https://www.zhihu.com/question/1891113677651955809/answer/1915210910441649247>) | B | 沿用既有阅读记录 |
| 197 | [Stripe Sessions 2026：Agentic commerce 与智能体支付基础设施](<https://stripe.com/blog/everything-we-announced-at-sessions-2026>) | B | 沿用既有阅读记录 |
| 198 | [胡渊鸣：生成式 AI 这门生意（上）](<https://mp.weixin.qq.com/s/Zzj8RJjhWsGWTAs-qTdNdA>) | B | 沿用既有阅读记录 |
| 199 | [Nicholas Wilt 推荐的计算机书单](<http://xhslink.com/o/1dFyzo3fpY8>) | B | 沿用既有阅读记录 |
| 200 | [小红书：为什么人类活不过二百岁](<http://xhslink.com/o/3HXfFQhJZYz>) | B | 沿用既有阅读记录 |
| 201 | [大模型第一性原理：信息论、定向信息与 token 信道抽象](<https://mp.weixin.qq.com/s/MxcJOocm_xOyoGigcL0G2Q>) | B | 沿用既有阅读记录 |
| 202 | [Artem Kirsanov：概率背后的核心方程，熵、交叉熵与 KL 散度](<https://www.youtube.com/watch?v=KHVR587oW8I>) | B | 沿用既有阅读记录 |
| 203 | [DataFunTalk：字节和腾讯 Agentic RL 最佳实践议题线索](<https://mp.weixin.qq.com/s/LHphRzY-Un5Rr7HpSy403g>) | B | 沿用既有阅读记录 |
| 204 | [大模型的第一性原理（一）：统计物理篇](<https://mp.weixin.qq.com/s/zj4yzt0fPSR8QRevW3_ehg>) | B | 沿用既有阅读记录 |
| 205 | [Visual Language Hypothesis：为什么自监督难以直接学到语义](<http://xhslink.com/o/5opijYbkC2d>) · [来源 2](<https://arxiv.org/abs/2512.23335>) | B | 沿用既有阅读记录 |
| 206 | [知乎专栏文章：标题暂不可见](<https://zhuanlan.zhihu.com/p/1914885010231625343>) | U | 未读 |
| 207 | [Textual SGD: Optimizing External Memory](<http://xhslink.com/o/1mQp3aTSBjg>) | U | 未读 |
| 208 | [Static AI to Continual Enterprise Learning: A Living Dialect of Tribal Knowledge](<https://scale.com/blog/dialect-of-tribal-knowledge>) | S | 沿用既有阅读记录 |
| 209 | [Sema Code: Decoupling AI Coding Agents from the Terminal](<https://arxiv.org/abs/2604.09226>) | S | 沿用既有阅读记录 |
| 210 | [Thinking with Visual Primitives](<https://github.com/deepseek-ai/Thinking-with-Visual-Primitives>) | A | 沿用既有阅读记录 |
| 211 | [Microsoft Agent 365 GA](<https://www.microsoft.com/en-us/security/blog/2026/05/01/microsoft-agent-365-is-now-generally-available-the-control-plane-for-agents/>) | A | 沿用既有阅读记录 |
| 212 | [Introducing Cloudflare Agent Cloud](<https://blog.cloudflare.com/introducing-cloudflare-agent-cloud/>) | A | 沿用既有阅读记录 |
| 213 | [Sema Code: Decoupling AI Coding Agents into Programmable, Embeddable Infrastructure](<https://arxiv.org/abs/2604.11045>) | S | 沿用既有阅读记录 |
| 214 | [MemEvoBench: Benchmarking Memory MisEvolution in LLM Agents](<https://arxiv.org/abs/2604.15774>) | S | 沿用既有阅读记录 |
| 215 | [Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers](<https://arxiv.org/abs/2603.07670>) | A | 沿用既有阅读记录 |
| 216 | [Joel Leibo 研究脉络：multi-agent 的对象不是单个 agent，而是关系结构](<http://xhslink.com/o/5758vppKO2O>) | A | 沿用既有阅读记录 |
| 217 | [Reinforcement Learning for LLM-based Multi-Agent Systems through Orchestration Traces](<https://arxiv-troller.com/paper/3162275/>) | S | 沿用既有阅读记录 |
| 218 | [From Skill Text to Skill Structure: The Scheduling-Structural-Logical Representation for Agent Skills](<https://papers.cool/arxiv/2604.24026>) | S | 沿用既有阅读记录 |
| 219 | [GPT-5.5 Instant: smarter, clearer, and more personalized](<https://openai.com/index/gpt-5-5-instant/>) | A | 沿用既有阅读记录 |
| 220 | [Gemini API File Search is now multimodal](<https://blog.google/innovation-and-ai/technology/developers-tools/expanded-gemini-api-file-search-multimodal-rag/>) | A | 沿用既有阅读记录 |
| 221 | [How OpenAI delivers low-latency voice AI at scale](<https://openai.com/index/delivering-low-latency-voice-ai-at-scale/>) | A | 沿用既有阅读记录 |
| 222 | [SpecKV: Adaptive Speculative Decoding with Compression-Aware Gamma Selection](<https://arxiv-troller.com/paper/3162362/>) | A | 沿用既有阅读记录 |
| 223 | [OpenAI MRC: Supercomputer networking to accelerate large scale AI training](<https://openai.com/index/mrc-supercomputer-networking/>) | S | 沿用既有阅读记录 |
| 224 | [Agents can now create Cloudflare accounts, buy domains, and deploy](<https://blog.cloudflare.com/agents-stripe-projects/>) | S | 沿用既有阅读记录 |
| 225 | [Anthropic: Higher usage limits for Claude and a compute deal with SpaceX](<https://www.anthropic.com/news/higher-limits-spacex>) | A | 沿用既有阅读记录 |
| 226 | [Gemini API File Search is now multimodal](<https://blog.google/innovation-and-ai/technology/developers-tools/expanded-gemini-api-file-search-multimodal-rag/>) | A | 沿用既有阅读记录 |
| 227 | [Gemini API Webhooks for long-running jobs](<https://blog.google/innovation-and-ai/technology/developers-tools/event-driven-webhooks/>) | A | 沿用既有阅读记录 |
| 228 | [CopilotKit raises $27M Series A](<https://www.copilotkit.ai/blog/series-a>) | A | 沿用既有阅读记录 |
| 229 | [Model-Harness-Fit：同一模型换壳性能会显著漂移](<https://mp.weixin.qq.com/s/TTe7IY_pjAuv4zA9krlYrQ>) | A | 沿用既有阅读记录 |
| 230 | [Anthropic: Natural Language Autoencoders: Turning Claude's Thoughts into Text](<https://www.anthropic.com/research/natural-language-autoencoders>) · [来源 2](<https://github.com/kitft/natural_language_autoencoders>) | S | 沿用既有阅读记录 |
| 231 | [Reinforcement Learning for LLM-based Multi-Agent Systems through Orchestration Traces](<https://huggingface.co/papers/2605.02801>) · [来源 2](<https://github.com/xxzcc/awesome-llm-mas-rl>) | S | 沿用既有阅读记录 |
| 232 | [Do Agent Rules Shape or Distort? Guardrails Beat Guidance in Coding Agents](<https://arxiv.org/abs/2604.11088>) | A | 沿用既有阅读记录 |
| 233 | [agent-skills-eval](<https://github.com/darkrishabh/agent-skills-eval>) | A | 沿用既有阅读记录 |
| 234 | [Accelerating Gemma 4: faster inference with multi-token prediction drafters](<https://blog.google/innovation-and-ai/technology/developers-tools/multi-token-prediction-gemma-4/>) | A | 沿用既有阅读记录 |
| 235 | [ds4: DeepSeek 4 Flash local inference engine for Metal](<https://github.com/antirez/ds4>) | A | 沿用既有阅读记录 |
| 236 | [OpenAI: Advancing voice intelligence with new models in the API](<https://openai.com/index/advancing-voice-intelligence-with-new-models-in-the-api/>) | A | 沿用既有阅读记录 |
| 237 | [AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields](<https://deepmind.google/blog/alphaevolve-impact/>) | A | 沿用既有阅读记录 |
| 238 | [NVIDIA Dynamo Agent Context and Tracing](<https://docs.nvidia.com/dynamo/dev/user-guides/agents/agent-context-and-tracing>) | A | 沿用既有阅读记录 |
| 239 | [Ray A2A multi-agent + Anyscale scalable MCP on Ray Serve](<https://docs.ray.io/en/latest/ray-overview/examples/multi_agent_a2a/content/README.html>) | A | 沿用既有阅读记录 |
| 240 | [OpenAI Codex browser / Chrome workflow signal](<https://openai.com/index/codex-for-almost-everything/>) | A | 沿用既有阅读记录 |
| 241 | [Claude Microsoft 365 connector and enterprise context permissions](<https://support.claude.com/en/articles/12542951-enable-and-use-the-microsoft-365-connector>) | A | 沿用既有阅读记录 |
| 242 | [STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?](<https://arxiv.org/abs/2605.06527>) | S | 沿用既有阅读记录 |
| 243 | [LLMs Corrupt Your Documents When You Delegate](<https://arxiv.org/abs/2604.15597>) | S | 沿用既有阅读记录 |
| 244 | [Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents](<https://arxiv.org/abs/2605.06635>) | S | 沿用既有阅读记录 |
| 245 | [SysMoBench / Specula: Can LLMs model real-world systems in TLA+?](<https://www.sigops.org/2026/can-llms-model-real-world-systems-in-tla/>) · [来源 2](<https://github.com/specula-org/SysMoBench>) | A | 沿用既有阅读记录 |
| 246 | [re_gent: Version-Control for AI coding agents](<https://github.com/regent-vcs/re_gent>) | A | 沿用既有阅读记录 |
| 247 | [A recent experience with ChatGPT 5.5 Pro](<https://gowers.wordpress.com/2026/05/08/a-recent-experience-with-chatgpt-5-5-pro/>) | A | 沿用既有阅读记录 |
| 248 | [MiniMax “马嘉祺”问题：后训练 token 覆盖不足导致输出映射层退化](<https://mp.weixin.qq.com/s/sBnue6tvfrtDU70UHiGcbQ>) | A | 沿用既有阅读记录 |
| 249 | [How AI-pilled are you? / P9 AI Fluency Index](<https://ai-pilled.com/>) | A | 沿用既有阅读记录 |
| 250 | [Jiayi Weng：Learning Beyond Gradients / Heuristic Learning](<https://trinkle23897.github.io/learning-beyond-gradients/>) | S | 沿用既有阅读记录 |
| 251 | [MiniMax：一个 AI 还是不够 / Agent Teams 与 Mavis](<https://mp.weixin.qq.com/s/TIL7o92f71DsPPLWT4_37A>) | A | 沿用既有阅读记录 |
| 252 | [甲子光年：Kimi 总裁张予彤北大实录：我们想要有抽象能力和偏执的人](<https://mp.weixin.qq.com/s/JrKeN3Gks7a5Hkr-b06UFw>) | A | 沿用既有阅读记录 |
| 253 | [MemReranker：Reasoning-Aware Reranking for Agent Memory Retrieval](<https://arxiv.org/abs/2605.06132>) | S | 沿用既有阅读记录 |
| 254 | [Thinking Machines：Interaction Models](<https://thinkingmachines.ai/blog/interaction-models/>) | S | 沿用既有阅读记录 |
| 255 | [OpenAI：Running Codex safely at OpenAI](<https://openai.com/index/running-codex-safely/>) | S | 沿用既有阅读记录 |
| 256 | [Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning](<https://arxiv.org/abs/2605.10923>) | S | 沿用既有阅读记录 |
| 257 | [WildClawBench：Real-World Long-Horizon Agent Evaluation](<https://arxiv.org/abs/2605.10912>) | A | 沿用既有阅读记录 |
| 258 | [Needle：26M function calling model](<https://github.com/cactus-compute/needle>) | A | 沿用既有阅读记录 |
| 259 | [Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering Tasks](<https://lab.einsia.ai/frontier-eng/>) · [来源 2](<https://huggingface.co/papers/2604.12290>) | S | 沿用既有阅读记录 |
| 260 | [OpenAI Daybreak / Codex Security control plane](<https://openai.com/daybreak>) | S | 沿用既有阅读记录 |
| 261 | [OpenAI: Investigating the consequences of accidentally grading CoT during RL](<https://alignment.openai.com/accidental-cot-grading/>) | A | 沿用既有阅读记录 |
| 262 | [UiPath for Coding Agents](<https://www.uipath.com/newsroom/uipath-for-coding-agents-launch>) | A | 沿用既有阅读记录 |
| 263 | [SGLang v0.5.11 release](<https://github.com/sgl-project/sglang/releases>) | A | 沿用既有阅读记录 |
| 264 | [ELF: Embedded Language Flows](<https://arxiv.org/abs/2605.10938>) · [来源 2](<https://github.com/lillian039/ELF>) | A | 沿用既有阅读记录 |
| 265 | [LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues](<https://arxiv.org/abs/2605.12493>) | S | 沿用既有阅读记录 |
| 266 | [TencentDB Agent Memory](<https://cloud.tencent.com/developer/article/2668579>) · [来源 2](<https://github.com/Tencent/TencentDB-Agent-Memory>) | S | 沿用既有阅读记录 |
| 267 | [Needle: 26M function call model](<https://github.com/cactus-compute/needle>) | A | 沿用既有阅读记录 |
| 268 | [Notion Developer Platform](<https://www.notion.com/blog/introducing-developer-platform>) | A | 沿用既有阅读记录 |
| 269 | [Claude Agent SDK credits and agent subscription economics](<https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan>) | A | 沿用既有阅读记录 |
| 270 | [Alexa for Shopping: personalized agentic commerce](<https://www.aboutamazon.com/news/retail/alexa-for-shopping-ai-assistant>) | A | 沿用既有阅读记录 |
| 271 | [Meta Incognito Chat with Meta AI](<https://about.fb.com/news/2026/05/incognito-chat-whatsapp-meta-ai/>) | B | 沿用既有阅读记录 |
| 272 | [Is Grep All You Need? How Agent Harnesses Reshape Agentic Search](<https://arxiv.org/abs/2605.15184>) | A | 沿用既有阅读记录 |
| 273 | [Qoder 1.0 / Quest Mode：AI IDE -> autonomous agent workbench](<https://developer.aliyun.com/article/1734787>) | A | 沿用既有阅读记录 |
| 274 | [ChatGPT Personal Finance](<https://openai.com/index/personal-finance-chatgpt/>) | B | 沿用既有阅读记录 |
| 275 | [Grok Build CLI](<https://x.ai/news/grok-build-cli>) | B | 沿用既有阅读记录 |
| 276 | [MeMo: Memory as a Model](<https://arxiv.org/abs/2605.15156>) | A | 沿用既有阅读记录 |
| 277 | [APWA: A Distributed Architecture for Parallelizable Agentic Workflows](<https://arxiv.org/abs/2605.15132>) | A | 沿用既有阅读记录 |
| 278 | [Ring-2.6-1T：面向 Agent 执行的开源推理模型](<https://huggingface.co/inclusionAI/Ring-2.6-1T>) | A | 沿用既有阅读记录 |
| 279 | [Anthropic: Teaching Claude why](<https://alignment.anthropic.com/2026/teaching-claude-why/>) | A | 沿用既有阅读记录 |
| 280 | [AGenUI：三端原生 A2UI Renderer](<https://github.com/AGenUI/AGenUI>) | B | 沿用既有阅读记录 |
| 281 | [TwiSTAR: Think Fast, Think Slow, Then Act](<https://arxiv.org/abs/2605.11553>) | A | 沿用既有阅读记录 |
| 282 | [小红书 paper radar：4月近两周自觉不错的 paper 合集第三期](<http://xhslink.com/o/44WZFV68pUu>) | A | 沿用既有阅读记录 |
| 283 | [ADAS / Automated Design of Agentic Systems：用搜索自动设计 agent system](<https://arxiv.org/abs/2408.08435>) · [来源 2](<https://github.com/ShengranHu/ADAS>) · [来源 3](<https://openreview.net/forum?id=t9U3LW7JVX>) | A | 沿用既有阅读记录 |
| 284 | [DGM / Darwin Gödel Machine：开放式自改进 agent](<https://arxiv.org/abs/2505.22954>) · [来源 2](<https://github.com/jennyzzt/dgm>) | A | 沿用既有阅读记录 |
| 285 | [AFlow：自动生成 agentic workflow](<https://arxiv.org/abs/2410.10762>) · [来源 2](<https://github.com/FoundationAgents/AFlow>) | A | 沿用既有阅读记录 |
| 286 | [SPO / Self-Supervised Prompt Optimization：无人工标签的 prompt 优化](<https://arxiv.org/abs/2502.06855>) · [来源 2](<https://github.com/FoundationAgents/SPO>) | A | 沿用既有阅读记录 |
| 287 | [GEPA / Reflective Prompt Evolution：反思式 prompt evolution](<https://arxiv.org/abs/2507.19457>) · [来源 2](<https://github.com/gepa-ai/gepa>) · [来源 3](<https://openreview.net/pdf?id=4oo6XTL6Oj>) | A | 沿用既有阅读记录 |
| 288 | [COMPASS：Enhancing Agent Long-Horizon Reasoning with Evolving Context](<https://arxiv.org/pdf/2510.08790>) | A | 沿用既有阅读记录 |
| 289 | [Training-Free Group Relative Policy Optimization](<https://arxiv.org/pdf/2510.08191>) | A | 沿用既有阅读记录 |
| 290 | [Motive Notes：What Makes 5% of AI Agents Actually Work in Production?](<https://www.motivenotes.ai/p/what-makes-5-of-ai-agents-actually>) | A | 沿用既有阅读记录 |
| 291 | [AgentEvolver：Towards Efficient Self-Evolving Agent System](<https://mp.weixin.qq.com/s/n8ZYMbs8yoKwX-eM73GgtQ>) · [来源 2](<https://arxiv.org/abs/2511.10395>) · [来源 3](<https://github.com/modelscope/AgentEvolver>) | A | 沿用既有阅读记录 |
| 292 | [To The Crazy Ones \| 致超级个体](<https://my.feishu.cn/wiki/AfvNwnZiEirKPSkmUz7cfDlYnM1>) | A | 沿用既有阅读记录；既往已读文字正文，视频、图片与白板未逐一展开。讨论完整问题所有权、工具权限、直接用户反馈与组织认可如何支持 AI Builder。 |
| 293 | [Codex / Claude Code memory 模式的收益、风险与 token 成本](<https://developers.openai.com/codex/memories>) · [来源 2](<https://arxiv.org/abs/2601.20404>) | A | 沿用既有阅读记录 |
| 294 | [SkyRL：full-stack RL library for LLMs](<https://docs.skyrl.ai/>) · [来源 2](<https://github.com/NovaSky-AI/SkyRL>) | A | 沿用既有阅读记录 |
| 295 | [OpenRLHF：high-performance RLHF / GRPO / PPO framework](<https://github.com/OpenRLHF/OpenRLHF>) | A | 沿用既有阅读记录 |
| 296 | [Selective Rollout：Efficient Reinforcement Learning for Long-Horizon LLM Agents](<https://arxiv.org/abs/2605.05802>) | A | 沿用既有阅读记录 |
| 297 | [HiPER：State Abstraction and Value-Guided Search for Efficient Long-Horizon Agents](<https://arxiv.org/abs/2602.16165>) | A | 沿用既有阅读记录 |
| 298 | [AgentFly：Extensible and Scalable Reinforcement Learning for LM Agents](<https://arxiv.org/abs/2507.14897>) | A | 沿用既有阅读记录 |
| 299 | [Code as Agent Harness](<https://huggingface.co/papers/2605.18747>) · [来源 2](<https://arxiv.org/abs/2605.18747>) | S | 沿用既有阅读记录 |
| 300 | [SWE-Chain：Benchmarking Coding Agents on Chained Release-Level Package Upgrades](<https://arxiv.org/abs/2605.14415>) | S | 沿用既有阅读记录 |
| 301 | [AgentTrust：Runtime Safety Evaluation and Interception for AI Agent Tool Use](<https://arxiv.org/abs/2605.04785>) | A | 沿用既有阅读记录 |
| 302 | [Google I/O 2026：Gemini 3.5 Flash / Antigravity 2.0 / Managed Agents](<https://blog.google/innovation-and-ai/technology/developers-tools/google-io-2026-developer-highlights/>) | A | 沿用既有阅读记录 |
| 303 | [OpenAI / Databricks：GPT-5.5 on OfficeQA Pro](<https://openai.com/index/databricks/>) | A | 沿用既有阅读记录 |
| 304 | [Anthropic acquires Stainless：SDK / CLI / MCP server tooling](<https://www.anthropic.com/news/anthropic-acquires-stainless>) | A | 沿用既有阅读记录 |
| 305 | [RRFP：A Readiness-Driven Runtime for Pipeline-Parallel Training under Runtime Variability](<https://arxiv.org/abs/2605.18750>) | A | 沿用既有阅读记录 |
| 306 | [DashAttention：Differentiable and Adaptive Sparse Hierarchical Attention](<https://arxiv.org/abs/2605.18753>) | A | 沿用既有阅读记录 |
| 307 | [Qwen3.7-Max Agent Frontier](<https://qwen.ai/blog?id=qwen3.7>) | A | 沿用既有阅读记录 |
| 308 | [OpenAI Guaranteed Capacity：模型 API 的 reserved compute contract](<https://openai.com/business/guaranteed-capacity/>) | A | 沿用既有阅读记录 |
| 309 | [TIDE：Efficient and Lossless MoE Diffusion LLM Inference with I/O-aware Expert Offload](<https://arxiv.org/abs/2605.20179>) | A | 沿用既有阅读记录 |
| 310 | [KoRe：Compact Knowledge Representations for Large Language Models](<https://arxiv.org/abs/2605.20170>) | A | 沿用既有阅读记录 |
| 311 | [OpenAI model disproves Erdős unit-distance conjecture](<https://openai.com/index/model-disproves-discrete-geometry-conjecture/>) | A | 沿用既有阅读记录 |
| 312 | [Multi-Stream LLMs：Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs](<https://arxiv.org/abs/2605.12460>) | A | 沿用既有阅读记录 |
| 313 | [vLLM x PegaFlow：Production-Grade External KV Cache](<https://vllm.ai/blog/2026-05-18-pegaflow>) | A | 沿用既有阅读记录 |
| 314 | [AWS SageMaker OpenAI-compatible API endpoints](<https://aws.amazon.com/blogs/machine-learning/announcing-openai-compatible-api-support-for-amazon-sagemaker-ai-endpoints/>) | A | 沿用既有阅读记录 |
| 315 | [Google AI Mode Ads：Gemini-powered ad formats in Search](<https://blog.google/products/ads-commerce/google-marketing-live-search-ads/>) | A | 沿用既有阅读记录 |
| 316 | [Tencent Hy-MT2：translation model family and IFMTBench](<https://github.com/Tencent-Hunyuan/Hy-MT2>) | A | 沿用既有阅读记录 |
| 317 | [Cohere Command A+：可私有部署的 open-source enterprise agent model](<https://cohere.com/blog/command-a-plus>) | A | 沿用既有阅读记录 |
| 318 | [Domain-Camouflaged Injection Attacks：领域伪装注入绕过 agent guard](<https://arxiv.org/abs/2605.22001>) | A | 沿用既有阅读记录 |
| 319 | [唐杰：关于 long-horizon tasks 的近期思考](<http://xhslink.com/o/5bvmo8slRJq>) | A | 沿用既有阅读记录 |
| 320 | [Runtime (YC P26)：团队级沙盒化 coding agents](<https://www.producthunt.com/products/runtime>) | B | 沿用既有阅读记录 |
| 321 | [Claude Code 源码精读：compact 每次压缩时发生了什么](<http://xhslink.com/o/3YETdBdSzxf>) | A | 沿用既有阅读记录 |
| 322 | [Neuromancer：Claude Code 上下文管理个人学习笔记](<http://xhslink.com/o/6PRR8PKYQ6o>) | A | 沿用既有阅读记录 |
| 323 | [MemGym：长程 agent memory 执行评测环境](<https://arxiv.org/abs/2605.20833>) | A | 沿用既有阅读记录 |
| 324 | [Anthropic Project Glasswing / Claude Mythos CVD dashboard](<https://www.anthropic.com/research/glasswing-initial-update>) | A | 沿用既有阅读记录 |
| 325 | [CODA：Rewriting Transformer Blocks as GEMM-Epilogue Programs](<https://arxiv.org/abs/2605.19269>) | A | 沿用既有阅读记录 |
| 326 | [KanBots OSS：本地 Kanban 多 agent 编排](<https://www.kanbots.dev/oss>) · [来源 2](<https://github.com/leodavinci1/kanbots>) | B | 沿用既有阅读记录 |
| 327 | [ReAct：reasoning-action-observation 循环的经典起点](<https://arxiv.org/abs/2210.03629>) | A | 沿用既有阅读记录 |
| 328 | [Toolformer：模型自监督学会调用工具](<https://arxiv.org/abs/2302.04761>) | A | 沿用既有阅读记录 |
| 329 | [API-Bank：tool-augmented LLM 的 API 评测早期基准](<https://arxiv.org/abs/2304.08244>) | A | 沿用既有阅读记录 |
| 330 | [Gorilla：面向海量 API 的 tool retrieval / function calling](<https://arxiv.org/abs/2305.15334>) · [来源 2](<https://github.com/ShishirPatil/gorilla>) | A | 沿用既有阅读记录 |
| 331 | [ToolLLM / ToolBench：大规模真实 API 的工具学习与评测](<https://arxiv.org/abs/2307.16789>) · [来源 2](<https://github.com/OpenBMB/ToolBench>) | A | 沿用既有阅读记录 |
| 332 | [WebArena：真实 Web 环境中的 autonomous agent benchmark](<https://arxiv.org/abs/2307.13854>) | A | 沿用既有阅读记录 |
| 333 | [VisualWebArena：多模态 Web agent 评测](<https://arxiv.org/abs/2401.13649>) | A | 沿用既有阅读记录 |
| 334 | [OSWorld：真实桌面环境中的 computer-use agent benchmark](<https://arxiv.org/abs/2404.07972>) · [来源 2](<https://github.com/xlang-ai/OSWorld>) | A | 沿用既有阅读记录 |
| 335 | [WorkArena：企业知识工作 Web agent benchmark](<https://arxiv.org/abs/2403.07718>) · [来源 2](<https://arxiv.org/abs/2407.05291>) | A | 沿用既有阅读记录 |
| 336 | [SWE-bench：真实 GitHub issue 到 patch 的软件工程评测](<https://arxiv.org/abs/2310.06770>) | A | 沿用既有阅读记录 |
| 337 | [OpenHands：通用软件开发 agent 平台](<https://arxiv.org/abs/2407.16741>) · [来源 2](<https://github.com/All-Hands-AI/OpenHands>) | A | 沿用既有阅读记录 |
| 338 | [AutoGen：multi-agent conversation framework](<https://arxiv.org/abs/2308.08155>) · [来源 2](<https://github.com/microsoft/autogen>) | A | 沿用既有阅读记录 |
| 339 | [LLMCompiler：并行 function calling / tool execution 编排](<https://arxiv.org/abs/2312.04511>) | A | 沿用既有阅读记录 |
| 340 | [AgentLens：agent 行为可视分析与 lucky pass 问题](<https://arxiv.org/abs/2402.08995>) · [来源 2](<https://arxiv.org/abs/2605.12925>) | A | 沿用既有阅读记录 |
| 341 | [Contextual Agent Security：面向不同目的的 agent policy](<https://arxiv.org/abs/2501.17070>) | A | 沿用既有阅读记录 |
| 342 | [Generative Agents：长期记忆驱动的交互式行为模拟](<https://arxiv.org/abs/2304.03442>) | A | 沿用既有阅读记录 |
| 343 | [Lost in the Middle：长上下文位置偏置经典问题](<https://arxiv.org/abs/2307.03172>) | A | 沿用既有阅读记录 |
| 344 | [MemGPT：把 LLM memory 管理类比为操作系统](<https://arxiv.org/abs/2310.08560>) · [来源 2](<https://github.com/letta-ai/letta>) | A | 沿用既有阅读记录 |
| 345 | [OpenShell：声明式 policy 驱动的 agent sandbox runtime](<https://github.com/NVIDIA/OpenShell>) | A | 沿用既有阅读记录 |
| 346 | [SWE-ReX：coding agent remote execution / sandbox infrastructure](<https://github.com/SWE-agent/SWE-ReX>) | A | 沿用既有阅读记录 |
| 347 | [ContextForge：MCP / A2A / REST gateway with governance and observability](<https://github.com/IBM/mcp-context-forge>) | A | 沿用既有阅读记录 |
| 348 | [Agent Governance Toolkit：deterministic policy / identity / sandbox / audit before actions](<https://github.com/microsoft/agent-governance-toolkit>) | A | 沿用既有阅读记录 |
| 349 | [Browser Harness：可编辑 CDP browser harness](<https://github.com/browser-use/browser-harness>) | A | 沿用既有阅读记录 |
| 350 | [Symphony：ticket-driven orchestration layer for autonomous implementation runs](<https://github.com/openai/symphony>) | A | 沿用既有阅读记录 |
| 351 | [R2E-Gym：从真实 repo issue 构造 executable coding-agent RL environments](<https://arxiv.org/abs/2504.07164>) · [来源 2](<https://github.com/R2E-Gym/R2E-Gym>) | A | 沿用既有阅读记录 |
| 352 | [Prime Intellect verifiers：LLM RL environments + evals as reusable verifier library](<https://github.com/PrimeIntellect-ai/verifiers>) | A | 沿用既有阅读记录 |
| 353 | [Meta-Harness：把 harness design 本身作为 automated search object](<https://openreview.net/forum?id=Qti7OhJr2P>) | A | 沿用既有阅读记录 |
| 354 | [Anthropic Context Management：tool result clearing and compaction](<https://www.anthropic.com/news/context-management>) | A | 沿用既有阅读记录 |
| 355 | [Context Rot：long context 变长时的性能退化](<https://research.trychroma.com/context-rot>) | A | 沿用既有阅读记录 |
| 356 | [Anthropic: How we built our multi-agent research system](<https://www.anthropic.com/engineering/multi-agent-research-system>) | A | 沿用既有阅读记录 |
| 357 | [LCGuard：Defending Against Latent Communication in Multi-Agent Systems by System-Level KV Cache Sandboxing](<https://arxiv.org/abs/2605.20920>) | A | 沿用既有阅读记录 |
| 358 | [Claude Code network sandbox bypass reports](<https://www.theregister.com/security/2026/05/20/even-claude-agrees-hole-in-its-sandbox-was-real-and-dangerous/5243662>) | A | 沿用既有阅读记录 |
| 359 | [Cloudflare Agent Infrastructure Stack](<https://www.infoq.com/news/2026/05/cloudflare-agent-platform-stack/>) | A | 沿用既有阅读记录 |
| 360 | [Reasonix：prefix-cache-aware terminal coding agent](<https://github.com/esengine/Reasonix>) | A | 沿用既有阅读记录 |
| 361 | [FAME：Fault-Aware Mixture-of-Experts for Message-Level Log Anomaly Detection](<https://arxiv.org/abs/2605.18504>) | A | 沿用既有阅读记录 |
| 362 | [Epoch AI: AI chip component cost shares](<https://epoch.ai/data-insights/ai-chip-component-cost-shares>) | A | 沿用既有阅读记录 |
| 363 | [Kung & Robinson: On Optimistic Methods for Concurrency Control](<https://www.eecs.harvard.edu/~htk/publication/1981-tods-kung-robinson.pdf>) | A | 沿用既有阅读记录 |
| 364 | [Shapiro et al.: A comprehensive study of convergent and commutative replicated data types（CRDTs）](<https://webarchive.di.uminho.pt/haslab.uminho.pt/cbm/publications/comprehensive-study-convergent-and-commutative-replicated-data-types.html>) | A | 沿用既有阅读记录 |
| 365 | [AeSlides：通过可验证奖励强化幻灯片生成](<http://xhslink.com/o/8UA38erc2GO>) · [来源 2](<https://arxiv.org/abs/2604.22840>) · [来源 3](<https://github.com/ympan0508/aeslides>) | A | 沿用既有阅读记录 |
| 366 | [ACC: Compiling Agent Trajectories for Long-Context Training](<https://arxiv.org/abs/2605.21850>) | A | 沿用既有阅读记录 |
| 367 | [SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?](<https://unipat.ai/benchmarks/SaaS-Bench>) · [来源 2](<https://arxiv.org/abs/2605.15777>) | A | 沿用既有阅读记录 |
| 368 | [π-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows](<https://simplified-reasoning.github.io/Pi-Bench/>) · [来源 2](<https://arxiv.org/abs/2605.14678>) | A | 沿用既有阅读记录 |
| 369 | [Microsoft Copilot Cowork Exfiltrates Files](<https://www.promptarmor.com/resources/microsoft-copilot-cowork-exfiltrates-files>) | A | 沿用既有阅读记录 |
| 370 | [Constraint Decay: The Fragility of LLM Agents in Backend Code Generation](<https://arxiv.org/abs/2605.06445>) | A | 沿用既有阅读记录 |
| 371 | [CVEvolve: Autonomous Algorithm Discovery for Unstructured Scientific Data Processing](<https://arxiv.org/abs/2605.11359>) | A | 沿用既有阅读记录 |
| 372 | [Gemini app becomes more agentic, delivering proactive 24/7 help](<https://blog.google/innovation-and-ai/products/gemini-app/next-evolution-gemini-app/>) | B | 沿用既有阅读记录 |
| 373 | [Anthropic: How we contain Claude across products](<https://www.anthropic.com/engineering/how-we-contain-claude>) | A | 沿用既有阅读记录 |
| 374 | [Language Models Need Sleep](<https://arxiv.org/abs/2605.26099>) | A | 沿用既有阅读记录 |
| 375 | [EAGLE 3.1: Advancing Speculative Decoding Through Collaboration Between EAGLE, vLLM, and TorchSpec](<https://vllm.ai/blog/2026-05-26-eagle-3-1>) | A | 沿用既有阅读记录 |
| 376 | [Robin: A multi-agent system for automating scientific discovery](<https://www.nature.com/articles/s41586-026-10652-y>) | A | 沿用既有阅读记录 |
| 377 | [Alipay AI Wallet / Token Pay / Agentic Commerce Trust Protocol](<https://www.businesswire.com/news/home/20260526337824/en/Alipay-Launches-Next-Generation-AI-Payment-Infrastructure-Debuts-AI-Wallet-and-Token-Pay-to-Power-Agentic-Economy>) | A | 沿用既有阅读记录 |
| 378 | [OpenRouter Raises $113M Series B](<https://www.businesswire.com/news/home/20260526953416/en/OpenRouter-Raises-%24113-Million-CapitalG-led-Series-B-as-Weekly-Volume-Explodes-to-25T-Tokens>) | A | 沿用既有阅读记录 |
| 379 | [Xiaomi MiMo-V2.5 Series Price Adjustment](<https://platform.xiaomimimo.com/docs/en-US/news/v2.5-price-update>) | A | 沿用既有阅读记录 |
| 380 | [Minicor: managed self-healing desktop automation at scale](<https://www.ycombinator.com/companies/minicor>) | A | 沿用既有阅读记录 |
| 381 | [中国企业家：6个月融25亿元，他是“字节系”最猛的AI创业者](<https://mp.weixin.qq.com/s/LPwU22G74CkFXZbuFGeetg>) | A | 沿用既有阅读记录 |
| 382 | [ArkClaw 漫剧虾工作流实测：从一个主题到爆款漫剧成片](<https://mp.weixin.qq.com/s/Tq3Igidp3PNhWvke_nyQfg>) | A | 沿用既有阅读记录 |
| 383 | [视频生成 agent / 短剧工具竞品池：Flova / 纳米短剧 / 巨日禄 / 万镜一刻](<https://www.flova.ai/zh-CN/projects/>) | A | 沿用既有阅读记录 |
| 384 | [火山引擎：Vibe Creating，让视频创作回归表达本身](<https://mp.weixin.qq.com/s/1q6hfza-EWKnNHAsIV9Ahg>) | A | 沿用既有阅读记录 |
| 385 | [MUSE-Autoskill：Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation](<http://xhslink.com/o/7cBqcdYJXWB>) · [来源 2](<https://arxiv.org/abs/2605.27366>) · [来源 3](<https://huggingface.co/papers/2605.27366>) | A | 沿用既有阅读记录 |
| 386 | [QUEST：Training Frontier Deep Research Agents with Fully Synthetic Tasks](<https://arxiv.org/abs/2605.24218>) | A | 沿用既有阅读记录 |
| 387 | [Cognition: More Devins in More Places](<https://cognition.ai/blog/series-d>) | A | 沿用既有阅读记录 |
| 388 | [jxnlco: Getting the most out of Codex](<https://x.com/jxnlco/status/2057153744630890620>) | A | 沿用既有阅读记录 |
| 389 | [Polar: Agentic RL on Any Harness at Scale](<https://arxiv.org/abs/2605.24220>) · [来源 2](<https://github.com/NVIDIA-NeMo/ProRL-Agent-Server>) | A | 沿用既有阅读记录 |
| 390 | [Structured Agent Distillation for Large Language Model](<https://arxiv.org/abs/2505.13820>) | A | 沿用既有阅读记录 |
| 391 | [PEFT-Arena: Understanding Parameter-Efficient Finetuning from a Stability-Plasticity Perspective](<https://arxiv.org/abs/2605.28819>) | A | 沿用既有阅读记录 |
| 392 | [Lenz Research: Beyond Benchmarks, Frontier LLM Disagreement on Fact-Checks](<https://lenz.io/research/llm-disagreement>) | A | 沿用既有阅读记录 |
| 393 | [DBOS: Postgres is All You Need for Durable Workflows](<https://www.dbos.dev/blog/postgres-is-all-you-need-for-durable-execution>) | A | 沿用既有阅读记录 |
| 394 | [OpenAI: How OpenAI uses Codex](<https://cdn.openai.com/pdf/6a2631dc-783e-479b-b1a4-af0cfbd38630/how-openai-uses-codex.pdf>) | A | 沿用既有阅读记录 |
| 395 | [ClickHouse Agents + Langfuse V4：agentic data stack and observability](<https://clickhouse.com/blog/open-house-2026-day-1>) | A | 沿用既有阅读记录 |
| 396 | [Claude Code vs Codex scientific-computing head-to-head](<https://arxiv.org/abs/2605.28916>) | A | 沿用既有阅读记录 |
| 397 | [Coding Beyond Your Training: Claude Code and the Technological Frontier of Software Developers](<https://arxiv.org/abs/2605.25438>) | A | 沿用既有阅读记录 |
| 398 | [SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?](<https://arxiv.org/abs/2605.30329>) | A | 沿用既有阅读记录 |
| 399 | [LLMSurgeon: Diagnosing Data Mixture of Large Language Models](<https://arxiv.org/abs/2605.30348>) | A | 沿用既有阅读记录 |
| 400 | [In-Context Reward Adaptation for Robust Preference Modeling](<https://arxiv.org/abs/2605.30323>) | A | 沿用既有阅读记录 |
| 401 | [Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players](<https://arxiv.org/abs/2605.28816>) | A | 沿用既有阅读记录 |
| 402 | [WALL-WM：World Action Model at Event Boundaries](<https://www.prnewswire.com/news-releases/x-square-robot-open-sources-wall-wm-shifting-robot-world-modeling-from-chunks-to-events-302785692.html>) · [来源 2](<https://github.com/X-Square-Robot/wall-x>) | A | 沿用既有阅读记录 |
| 403 | [ATLAS - Autoformalized Textbook Library At Scale](<https://github.com/facebookresearch/atlas-lean>) | A | 沿用既有阅读记录 |
| 404 | [tiny-vLLM：Build your own high performance LLM inference engine in C++ and CUDA](<https://github.com/jmaczan/tiny-vllm>) | A | 沿用既有阅读记录 |
| 405 | [Never Stop Learning: Continual Learning and Self-Iteration in LLMs](<https://victorchen96.github.io/continual_learning_survey.pdf>) | A | 沿用既有阅读记录 |
| 406 | [ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents](<https://arxiv.org/abs/2605.12481>) | A | 沿用既有阅读记录 |
| 407 | [Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory](<https://arxiv.org/abs/2605.26252>) | A | 沿用既有阅读记录 |
| 408 | [GodeX：OpenAI Responses API 兼容网关与 provider bridge](<https://github.com/Ahoo-Wang/GodeX>) | A | 沿用既有阅读记录 |
| 409 | [Meta AI support / Instagram account recovery exploit report](<https://www.0xsid.com/blog/meta-account-takeover-fiasco>) | A | 沿用既有阅读记录 |
| 410 | [Microsoft MAI / Frontier Tuning：workflow-specific RLE 与 Copilot harness model](<https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/>) | A | 沿用既有阅读记录 |
| 411 | [OpenAI Codex for every role / Sites / annotations](<https://openai.com/index/codex-for-every-role-tool-workflow/>) | A | 沿用既有阅读记录 |
| 412 | [Microsoft Scout / WorkIQ / Agent 365：always-on work agent control plane](<https://blogs.microsoft.com/blog/2026/06/02/microsoft-build-2026-be-yourself-at-work/>) | A | 沿用既有阅读记录 |
| 413 | [Bernini: Latent Semantic Planning for Video Diffusion](<https://github.com/bytedance/Bernini>) | A | 沿用既有阅读记录 |
| 414 | [AdaCodec: A Predictive Visual Code for Video MLLMs](<https://arxiv.org/abs/2606.02569>) | A | 沿用既有阅读记录 |
| 415 | [CLI-Anything: Towards Agent-Native Computer Use](<https://arxiv.org/abs/2606.03854>) · [来源 2](<https://github.com/HKUDS/CLI-Anything>) | A | 沿用既有阅读记录 |
| 416 | [EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Management](<https://arxiv.org/abs/2606.03841>) · [来源 2](<https://github.com/usail-hkust/EvoDS>) | A | 沿用既有阅读记录 |
| 417 | [Taiji: Pareto Optimal Policy Optimization with Semantics-IDs Trade-off for Industrial LLM-Enhanced Recommendation](<https://arxiv.org/abs/2606.03866>) | A | 沿用既有阅读记录 |
| 418 | [Bringing up DeepSeek-V4-Flash on AMD MI300X](<https://fergusfinn.com/blog/deepseek-v4-flash-mi300x/>) · [来源 2](<https://github.com/doublewordai/vllm-amd-blog-doubleword>) | B | 沿用既有阅读记录 |
| 419 | [Uber AI coding budget cap / enterprise agent FinOps](<https://news.bloomberglaw.com/artificial-intelligence/uber-caps-usage-of-ai-tools-like-claude-code-to-cut-costs-1>) | B | 沿用既有阅读记录 |
| 420 | [MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents](<https://arxiv.org/abs/2605.03952>) | A | 沿用既有阅读记录 |
| 421 | [KVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning Tasks](<https://arxiv.org/abs/2606.03458>) · [来源 2](<https://github.com/huawei-csl/KVarN>) | A | 沿用既有阅读记录 |
| 422 | [Multi-Segment Attention / AsymCache：面向 agent serving 的 KV-cache 管理](<https://arxiv.org/abs/2606.02964>) | A | 沿用既有阅读记录 |
| 423 | [Google Gemma 4 12B: unified encoder-free multimodal model](<https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12B/>) · [来源 2](<https://huggingface.co/google/gemma-4-12B>) | A | 沿用既有阅读记录 |
| 424 | [Anthropic / Sakana AI recursive self-improvement signals](<https://www.anthropic.com/institute/recursive-self-improvement>) | A | 沿用既有阅读记录 |
| 425 | [Microsoft pg_durable: PostgreSQL in-database durable execution](<https://github.com/microsoft/pg_durable>) | A | 沿用既有阅读记录 |
| 426 | [Anthropic Defending Code Reference Harness](<https://github.com/anthropics/defending-code-reference-harness>) | A | 沿用既有阅读记录 |
| 427 | [Alibaba Open Code Review: deterministic engineering + LLM agent code review](<https://github.com/alibaba/open-code-review>) | A | 沿用既有阅读记录 |
| 428 | [Cloudflare AI Gateway spend limits / identity-driven budgets](<https://blog.cloudflare.com/ai-gateway-spend-limits/>) | B | 沿用既有阅读记录 |
| 429 | [Lowfat: local CLI output filtering for agent token budgets](<https://github.com/zdk/lowfat>) | B | 沿用既有阅读记录 |
| 430 | [AdaMEM: Test-Time Adaptive Memory for Language Agents](<https://arxiv.org/abs/2606.05684>) · [来源 2](<https://github.com/yunx-z/AdaMEM>) | A | 沿用既有阅读记录 |
| 431 | [Vortex: Efficient and Programmable Sparse Attention Serving for AI Agents](<https://arxiv.org/abs/2606.06453>) | A | 沿用既有阅读记录 |
| 432 | [Description-Code Inconsistency in Real-world MCP Servers](<https://arxiv.org/abs/2606.04769>) | A | 沿用既有阅读记录 |
| 433 | [MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery](<https://arxiv.org/abs/2606.06473>) · [来源 2](<https://github.com/InternScience/MLEvolve>) | A | 沿用既有阅读记录 |
| 434 | [CollabSim: A CSCW-Grounded Methodology for Investigating Collaborative Competence of LLM Agents through Controlled Multi-Agent Experiments](<https://arxiv.org/abs/2606.06399>) | A | 沿用既有阅读记录 |
| 435 | [Scaffold, Not Vocabulary? A Controlled, Two-Tier, Pre-Registered Study of a Popperian Code-Generation Skill](<https://arxiv.org/abs/2606.06454>) | A | 沿用既有阅读记录 |
| 436 | [Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators](<https://arxiv.org/abs/2606.06476>) | A | 沿用既有阅读记录 |
| 437 | [TechCrunch: The token bill comes due](<https://techcrunch.com/2026/06/05/the-token-bill-comes-due-inside-the-industry-scramble-to-manage-ais-runaway-costs/>) | B | 沿用既有阅读记录 |
| 438 | [Poke becomes the first AI agent on Apple Messages for Business](<https://techcrunch.com/2026/06/04/apple-approves-poke-as-the-first-ai-agent-on-its-messages-for-business-platform/>) | B | 沿用既有阅读记录 |
| 439 | [知乎回答：王导是也缩略《置身钉内》与钉钉 ONE 项目复盘](<https://www.zhihu.com/question/2046261330911482494/answer/2046318885889127259>) | B | 沿用既有阅读记录 |
| 440 | [小红书：学习如何从 0 训练一个 SOTA LLM](<http://xhslink.com/o/7balNZMreoo>) | B | 沿用既有阅读记录 |
| 441 | [Motus: A Unified Latent Action World Model](<https://arxiv.org/abs/2512.13030>) | A | 沿用既有阅读记录 |
| 442 | [AKO: Agentic Kernel Optimization / AKO4ALL / AKO4X](<https://zhuanlan.zhihu.com/p/2044554843948308001>) · [来源 2](<https://github.com/TongmingLAIC/AKO4ALL>) · [来源 3](<https://github.com/TongmingLAIC/AKO4X>) | S | 沿用既有阅读记录 |
| 443 | [Nano World Models: minimalist world-model experiment substrate](<http://xhslink.com/o/2JbM85tWLmI>) · [来源 2](<https://arxiv.org/abs/2605.23993>) · [来源 3](<https://github.com/simchowitzlabpublic/nano-world-model>) | A | 沿用既有阅读记录 |
| 444 | [OpenAI Lockdown Mode：prompt injection 数据外泄防线的产品化 capability gate](<https://help.openai.com/en/articles/20001061-lockdown-mode>) | A | 沿用既有阅读记录 |
| 445 | [Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads](<https://arxiv.org/abs/2606.06448>) | A | 沿用既有阅读记录 |
| 446 | [SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents](<https://arxiv.org/abs/2606.05761>) · [来源 2](<https://github.com/KDEGroup/SubtleMemory>) | A | 沿用既有阅读记录 |
| 447 | [TIDE: Proactive Multi-Problem Discovery via Template-Guided Iteration](<https://arxiv.org/abs/2606.04743>) · [来源 2](<https://github.com/snap-stanford/tide>) | A | 沿用既有阅读记录 |
| 448 | [Goedel-Architect: Streamlining Formal Theorem Proving with Blueprint Generation and Refinement](<https://arxiv.org/abs/2606.06468>) | A | 沿用既有阅读记录 |
| 449 | [Google / SpaceX AI compute deal：短期 bridge capacity 与 frontier agent demand 的市场信号](<https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/>) | B | 沿用既有阅读记录 |
| 450 | [Cloudflare Bot Traffic Radar：agentic web traffic 与 origin cost 进入一等指标](<https://radar.cloudflare.com/bots>) | B | 沿用既有阅读记录 |
| 451 | [Tencent Productivity Agent Suite / CodeBuddy / WorkBuddy / Agent Runtime / TokenHub](<https://www.tencent.com/en-us/articles/2202350.html>) | B | 沿用既有阅读记录 |
| 452 | [FrontierCode：从 correctness 到 production mergeability 的 coding-agent benchmark](<https://cognition.ai/blog/frontier-code>) | A | 沿用既有阅读记录 |
| 453 | [KV cache serving exactness：Speculative KV coding + VeriCache](<https://fergusfinn.com/blog/kv-entropy-coder/>) · [来源 2](<https://arxiv.org/abs/2605.17613>) | A | 沿用既有阅读记录 |
| 454 | [Tokenomics / token bill：agentic software cost observability](<https://techcrunch.com/2026/06/05/the-token-bill-comes-due-inside-the-industry-scramble-to-manage-ais-runaway-costs/>) · [来源 2](<https://arxiv.org/abs/2601.14470>) | A | 沿用既有阅读记录 |
| 455 | [Intuned Agent：browser automation codegen + managed Playwright runtime](<https://intunedhq.com/>) | B | 沿用既有阅读记录 |
| 456 | [Microsoft AI developer tooling supply-chain incident](<https://techcrunch.com/2026/06/08/microsofts-open-source-tools-were-hacked-to-steal-passwords-of-ai-developers/>) | B | 沿用既有阅读记录 |
| 457 | [AGENTS.md / context-file tooling：agent-md-bench + context file evidence](<https://bernstein.run/tools/agent-md-bench>) | B | 沿用既有阅读记录 |
| 458 | [SWE-Explore：repository exploration benchmark for coding agents](<https://huggingface.co/papers/2606.07297>) · [来源 2](<https://arxiv.org/abs/2606.07297>) | A | 沿用既有阅读记录 |
| 459 | [End-to-End Context Compression at Scale / LCLM](<https://huggingface.co/papers/2606.09659>) · [来源 2](<https://arxiv.org/abs/2606.09659>) | A | 沿用既有阅读记录 |
| 460 | [Anthropic biology agents：domain data infrastructure for agents](<https://www.anthropic.com/research/agents-in-biology>) | A | 沿用既有阅读记录 |
| 461 | [Claude Fable 5 / Mythos 5：frontier capability with gated access](<https://www.anthropic.com/news/claude-fable-5-mythos-5>) | B | 沿用既有阅读记录 |
| 462 | [Nemotron 3 Ultra serving stack：vLLM / SGLang / Miles day-zero path](<https://vllm.ai/blog/2026-06-04-nemotron-3-ultra-vllm>) | B | 沿用既有阅读记录 |
| 463 | [AI developer tooling security：Microsoft repo incident + SGLang RCE](<https://techcrunch.com/2026/06/08/microsofts-open-source-tools-were-hacked-to-steal-passwords-of-ai-developers/>) · [来源 2](<https://github.com/advisories/GHSA-36m8-w8qf-g76p>) | B | 沿用既有阅读记录 |
| 464 | [Asuka-Bench：underspecified intent + multi-round refinement for code agents](<https://arxiv.org/abs/2606.05920>) · [来源 2](<https://arxiv.org/html/2606.05920>) | A | 沿用既有阅读记录 |
| 465 | [DiffusionGemma：parallel text diffusion for local interactive workflows](<https://developers.googleblog.com/diffusiongemma-the-developer-guide/>) | A | 沿用既有阅读记录 |
| 466 | [GitHub Agent Apps + Copilot Code Review skills/MCP](<https://github.blog/changelog/2026-06-02-extend-github-with-agent-apps/>) | A | 沿用既有阅读记录 |
| 467 | [MusaCoder：native GPU kernel generation with full-stack training on Moore Threads GPU](<https://arxiv.org/abs/2606.04847>) · [来源 2](<https://huggingface.co/MooreThreads/MusaCoder-27B>) | A | 沿用既有阅读记录 |
| 468 | [Apache Burr：state machine / telemetry / persistence for reliable AI apps](<https://github.com/apache/burr>) | A | 沿用既有阅读记录 |
| 469 | [Memory tools can make AI models worse：memory reliability as product risk](<https://techcrunch.com/2026/06/10/how-memory-tools-can-make-ai-models-worse/>) · [来源 2](<https://openreview.net/pdf?id=0Xt1qZ5xdW>) | B | 沿用既有阅读记录 |
| 470 | [MiMo Code：long-horizon coding agent with persistent project memory](<https://mimo.xiaomi.com/blog/mimo-code-long-horizon>) · [来源 2](<https://github.com/XiaomiMiMo/MiMo-Code>) | A | 沿用既有阅读记录 |
| 471 | [Claw Patrol：wire-level security firewall for agents](<https://github.com/denoland/clawpatrol>) | A | 沿用既有阅读记录 |
| 472 | [Google DeepMind multi-agent AI safety research fund](<https://deepmind.google/blog/investing-in-multi-agent-ai-safety-research/>) | A | 沿用既有阅读记录 |
| 473 | [Coinbase for Agents / x402：agentic payments and paid resource access](<https://www.coinbase.com/blog/coinbase-for-agents>) | A | 沿用既有阅读记录 |
| 474 | [Alibaba Cloud Meoo CLI：local coding agent to cloud deployment bridge](<https://news.aibase.com/news/28860>) | B | 沿用既有阅读记录 |
| 475 | [Can I Buy Your KV Cache?：agent-native prefill CDN / hot context cost model](<https://arxiv.org/abs/2606.13361>) | A | 沿用既有阅读记录 |
| 476 | [ReSum：self-summary as policy action for long reasoning RLVR](<https://arxiv.org/abs/2606.13316>) | A | 沿用既有阅读记录 |
| 477 | [AgentBeats：agentified agent assessment via A2A / MCP](<https://arxiv.org/abs/2606.13608>) | A | 沿用既有阅读记录 |
| 478 | [Agents-K1：agent-native scientific knowledge orchestration](<https://arxiv.org/abs/2606.13669>) · [来源 2](<https://arxiv.org/html/2606.13669v1>) · [来源 3](<https://github.com/InternScience/GraphAnything>) · [来源 4](<https://github.com/InternScience/GraphAnything/blob/abbe35dc12e4991181ff8c4eed0c1cd334ffad1b/README.md>) · [来源 5](<https://huggingface.co/InternScience/Agents-K1>) · [来源 6](<https://huggingface.co/datasets/InternScience/Scholar-kg>) | A | 沿用既有阅读记录 |
| 479 | [EurekAgent：environment engineering for autonomous scientific discovery](<https://arxiv.org/abs/2606.13662>) | A | 沿用既有阅读记录 |
| 480 | [SkillSpector：agent skill supply-chain scanner](<https://github.com/NVIDIA/SkillSpector>) | A | 沿用既有阅读记录 |
| 481 | [GLM-5：from vibe coding to agentic engineering](<https://arxiv.org/abs/2602.15763>) · [来源 2](<https://arxiv.org/html/2602.15763v1>) · [来源 3](<https://github.com/zai-org/GLM-5>) · [来源 4](<https://huggingface.co/zai-org/GLM-5>) | A | 沿用既有阅读记录 |
| 482 | [LongTraceRL：learning long-context reasoning from search-agent trajectories](<https://arxiv.org/abs/2605.31584>) · [来源 2](<https://arxiv.org/html/2605.31584v1>) · [来源 3](<https://github.com/THU-KEG/LongTraceRL>) | A | 沿用既有阅读记录 |
| 483 | [Plan-RewardBench / VPR：trajectory-level reward modeling and verifiable process reward for agents](<https://arxiv.org/abs/2604.08178>) · [来源 2](<https://arxiv.org/html/2604.08178v1>) · [来源 3](<https://huggingface.co/datasets/wyy1112/Plan-RewardBench>) · [来源 4](<https://arxiv.org/abs/2605.10325>) · [来源 5](<https://arxiv.org/html/2605.10325v1>) | A | 沿用既有阅读记录 |
| 484 | [EvoArena / EvoMem：tracking memory evolution for robust LLM agents](<https://arxiv.org/abs/2606.13681>) · [来源 2](<https://arxiv.org/html/2606.13681v1>) · [来源 3](<https://huggingface.co/papers/2606.13681>) | A | 沿用既有阅读记录 |
| 485 | [depthfirst 21 FFmpeg zero-days：autonomous security agent with reproducible PoCs](<https://depthfirst.com/research/21-zero-days-in-ffmpeg>) | A | 沿用既有阅读记录 |
| 486 | [OpenAI Codex enterprise workflow cases：Notion / Nextdoor / Wasmer outcome engineering](<https://openai.com/index/notion/>) | A | 沿用既有阅读记录 |
| 487 | [Microsoft Discovery GA：governed agentic R&D workflows](<https://azure.microsoft.com/en-us/blog/announcing-microsoft-discovery-general-availability-and-microsoft-discovery-app-preview/>) | A | 沿用既有阅读记录 |
| 488 | [TensorZero archive signal：LLMOps open-source continuity risk](<https://github.com/tensorzero>) · [来源 2](<https://github.com/tensorzero/tensorzero>) | B | 沿用既有阅读记录 |
| 489 | [OpenAI multistate investigation：AI product safety and personalization audit risk](<https://techcrunch.com/2026/06/13/openai-faces-investigation-from-state-attorneys-general/>) | B | 沿用既有阅读记录 |
| 490 | [WeaveBench：hybrid-interface long-horizon computer-use agent benchmark](<https://huggingface.co/papers/2606.09426>) · [来源 2](<https://arxiv.org/abs/2606.09426>) | A | 沿用既有阅读记录 |
| 491 | [TRACE：compiling user corrections into runtime enforcement for coding agents](<https://huggingface.co/papers/2606.13174>) · [来源 2](<https://arxiv.org/abs/2606.13174>) · [来源 3](<https://github.com/YujunZhou/tellonce>) | A | 沿用既有阅读记录 |
| 492 | [HarnessBridge：learnable bidirectional controller for LLM agent harness](<https://huggingface.co/papers/2606.12882>) · [来源 2](<https://arxiv.org/abs/2606.12882>) | A | 沿用既有阅读记录 |
| 493 | [EvoBrowseComp：benchmarking search agents on evolving knowledge](<https://huggingface.co/papers/2606.13120>) · [来源 2](<https://arxiv.org/abs/2606.13120>) | A | 沿用既有阅读记录 |
| 494 | [Perplexity / HBS：How AI Agents Reshape Knowledge Work](<https://research.perplexity.ai/articles/how-ai-agents-reshape-knowledge-work>) | A | 沿用既有阅读记录 |
| 495 | [Google DeepMind From AGI to ASI：multi-agent collectives as one ASI pathway](<https://deepmind.google/research/publications/239142/>) · [来源 2](<https://arxiv.org/abs/2606.12938>) | B | 沿用既有阅读记录 |
| 496 | [OpenAI Partner Network：enterprise AI delivery and specialization market](<https://openai.com/index/introducing-openai-partner-network/>) | B | 沿用既有阅读记录 |
| 497 | [Context window budget：Don't trust large context windows](<https://garrit.xyz/posts/2026-05-06-dont-trust-large-context-windows>) | B | 沿用既有阅读记录 |
| 498 | [AI provenance failure signal：UK police fake-evidence allegation + KPMG hallucinated report](<https://news.slashdot.org/story/26/06/14/0317211/uk-police-officer-accused-of-using-ai-to-fake-evidence>) | B | 沿用既有阅读记录 |
| 499 | [Gabriel Weinberg：No, everyone is not using AI for everything](<https://gabrielweinberg.com/p/people-are-consuming-ai-like-they>) | B | 沿用既有阅读记录 |
| 500 | [HarnessX：composable, adaptive, evolvable agent harness foundry](<https://arxiv.org/abs/2606.14249>) | A | 沿用既有阅读记录 |
| 501 | [StreamMemBench：streaming evaluation of agent memory for future-oriented assistance](<https://arxiv.org/abs/2606.14571>) · [来源 2](<https://github.com/landian60/StreamMemBench>) | A | 沿用既有阅读记录 |
| 502 | [Dialogue SWE-Bench：benchmarking dialogue-driven coding agents](<https://arxiv.org/abs/2606.13995>) · [来源 2](<https://arxiv.org/html/2606.13995v1>) | A | 沿用既有阅读记录 |
| 503 | [Parallel-Synthesis：direct latent-space synthesis for parallel branches in LLM-agent workflows](<https://arxiv.org/abs/2606.14672>) | A | 沿用既有阅读记录 |
| 504 | [SIMMER：latent failures in LLM executable planning with a world model](<https://arxiv.org/abs/2606.14574>) · [来源 2](<https://arxiv.org/html/2606.14574v1>) | A | 沿用既有阅读记录 |
| 505 | [Google Cloud OKF：Open Knowledge Format for agent-readable context bundles](<https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing>) · [来源 2](<https://github.com/GoogleCloudPlatform/knowledge-catalog>) | A | 沿用既有阅读记录 |
| 506 | [OpenRouter Fusion：model panels as an API-level reasoning primitive](<https://openrouter.ai/blog/announcements/fusion-beats-frontier/>) | A | 沿用既有阅读记录 |
| 507 | [Apple Foundation Models / Xcode 27：native provider protocol and agentic coding workflow](<https://www.apple.com/newsroom/2026/06/apple-aids-app-development-with-new-intelligence-frameworks-and-advanced-tools/>) | A | 沿用既有阅读记录 |
| 508 | [Enterprise agent identity and service-agent consolidation：NewCore + Salesforce / Fin](<https://techcrunch.com/2026/06/15/ai-agents-are-becoming-employees-newcore-emerges-with-66m-to-give-them-identities/>) | B | 沿用既有阅读记录 |
| 509 | [Measuring Agents in Production：真实生产 agent 仍是高频人工介入系统](<https://openreview.net/forum?id=FxNCt9xtOZ>) | A | 沿用既有阅读记录 |
| 510 | [Principles of Mixed-Initiative User Interfaces：混合主动权的经典设计原则](<https://dl.acm.org/doi/10.1145/302979.303030>) | A | 沿用既有阅读记录 |
| 511 | [Power to the People：Interactive ML 中人的角色](<https://www.microsoft.com/en-us/research/publication/power-to-the-people-the-role-of-humans-in-interactive-machine-learning/>) | A | 沿用既有阅读记录 |
| 512 | [Evaluation of Interactive Machine Learning Systems：algorithm-centered + human-centered 双验证](<https://arxiv.org/abs/1801.07964>) | A | 沿用既有阅读记录 |
| 513 | [A Benchmark for Scalable Oversight Mechanisms：监督机制也需要 benchmark](<https://openreview.net/forum?id=ePE9BMoh8L>) | A | 沿用既有阅读记录 |
| 514 | [Deep Reinforcement Learning from Human Preferences：少量偏好监督如何塑造复杂目标](<https://arxiv.org/abs/1706.03741>) | A | 沿用既有阅读记录 |
| 515 | [EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments](<https://arxiv.org/abs/2606.13681>) | S | 沿用既有阅读记录 |
| 516 | [Dialogue SWE-Bench: A Benchmark for Dialogue-Driven Coding Agents](<https://arxiv.org/abs/2606.13995>) | S | 沿用既有阅读记录 |
| 517 | [OpenRouter Fusion：多模型 deliberation 作为 API runtime primitive](<https://openrouter.ai/openrouter/fusion>) | A | 沿用既有阅读记录 |
| 518 | [OpenAI Deployment Simulation：用真实分布预演模型上线行为](<https://openai.com/index/deployment-simulation/>) | S | 沿用既有阅读记录 |
| 519 | [AA-AgentPerf：agentic inference 的 SLO / agents-per-megawatt 口径](<https://artificialanalysis.ai/articles/aa-agentperf>) | A | 沿用既有阅读记录 |
| 520 | [GLM-5.2：open-weight long-horizon agent model](<https://z.ai/blog/glm-5.2>) · [来源 2](<https://huggingface.co/zai-org/GLM-5.2>) | A | 沿用既有阅读记录 |
| 521 | [AI Coding Agents Can Reproduce Social Science Findings / SocSci-Repro-Bench](<https://arxiv.org/abs/2606.11447>) | A | 沿用既有阅读记录 |
| 522 | [LifeSciBench + AI Chemist：science agent 的专家 rubric 与湿实验闭环](<https://openai.com/index/introducing-life-sci-bench/>) | A | 沿用既有阅读记录 |
| 523 | [Appia / Pramaana：AI trust 从 policy 走向 conformity + proof](<https://www.linuxfoundation.org/press/linux-foundation-launches-appia-foundation-to-establish-standardized-conformity-specifications-across-the-ai-value-chain>) | A | 沿用既有阅读记录 |
| 524 | [MCP Enterprise-Managed Authorization：Zero-touch OAuth for MCP](<https://blog.modelcontextprotocol.io/posts/enterprise-managed-auth/>) | A | 沿用既有阅读记录 |
| 525 | [Elastic agent memory：hybrid retrieval + DLS 的生产 memory 参考](<https://www.elastic.co/search-labs/blog/agent-memory-elasticsearch>) | A | 沿用既有阅读记录 |
| 526 | [Decoupled Search Grounding：把 agent search 变成 MCP-compatible gateway](<https://arxiv.org/list/cs.AI/new>) | A | 仅元信息；原记录仅有聚合入口；未补齐具体原文，不作为已读材料。 |
| 527 | [RODS：multi-turn tool-use agent 的 reward-driven online data synthesis](<https://arxiv.org/list/cs.AI/new>) | A | 仅元信息；原记录仅有聚合入口；未补齐具体原文，不作为已读材料。 |
| 528 | [CEO-Bench：long-horizon business agent eval](<https://arxiv.org/list/cs.AI/new>) | A | 仅元信息；原记录仅有聚合入口；未补齐具体原文，不作为已读材料。 |
| 529 | [SGCD：GUI agent 的 off-trajectory continuation distillation](<https://arxiv.org/list/cs.AI/new>) | A | 仅元信息；原记录仅有聚合入口；未补齐具体原文，不作为已读材料。 |
| 530 | [Xcientist：AI scientist 的 research harness 与 claim drift 防线](<https://arxiv.org/list/cs.AI/new>) | A | 仅元信息；原记录仅有聚合入口；未补齐具体原文，不作为已读材料。 |
| 531 | [Stanford PhD 回山东做传统企业 AI 落地：代码快，诊断慢](<http://xhslink.com/o/tKqE2OASU8>) | A | 沿用既有阅读记录 |
| 532 | [Google Agentic Resource Discovery：agent capability discovery + trust manifest](<https://developers.googleblog.com/en/announcing-the-agentic-resource-discovery-specification/>) | A | 沿用既有阅读记录 |
| 533 | [Claude Design + Claude Code sync：design-to-code workspace 进入真实组件回路](<https://support.claude.com/en/articles/14604416-get-started-with-claude-design>) | A | 沿用既有阅读记录 |
| 534 | [SkillVetBench：LLM agent skills 的语义风险评估](<https://arxiv.org/abs/2606.15899>) | A | 沿用既有阅读记录 |
| 535 | [ADK Arena：用 LLM-as-a-Developer 测 agent framework 可用性](<https://arxiv.org/abs/2606.05548>) | A | 沿用既有阅读记录 |
| 536 | [Agent Planning Benchmark：把 planning failure 从执行失败里拆出来](<https://arxiv.org/abs/2606.04874>) | A | 沿用既有阅读记录 |
| 537 | [R3-Skill：skill routing 中的 rejection signal 不是垃圾数据](<https://arxiv.org/abs/2606.03565>) | A | 沿用既有阅读记录 |
| 538 | [Exploration Structure in LLM Agents：coding agent 的 repo traversal 结构会决定定位质量](<https://arxiv.org/abs/2606.11976>) | A | 沿用既有阅读记录 |
| 539 | [AgentFairBench：agent 公平性要测 action，不只测 answer](<https://arxiv.org/abs/2606.16723>) | A | 沿用既有阅读记录 |
| 540 | [Kimi Work / Kimi Code Goal Mode：本地长程 agent 的 goal state 与权限面](<https://www.kimi.com/resources/kimi-work-introduction>) | A | 沿用既有阅读记录 |
| 541 | [OpenAI Patch the Planet：安全 agent 的 patch loop 与 maintainer agency](<https://openai.com/index/patch-the-planet/>) | A | 沿用既有阅读记录 |
| 542 | [Google Interactions API GA：managed agent API contract](<https://blog.google/innovation-and-ai/technology/developers-tools/interactions-api-general-availability/>) | A | 沿用既有阅读记录 |
| 543 | [Multi-LCB + Contagion Networks：coding agent 评测的语言轴与 judge topology](<https://arxiv.org/abs/2606.20517>) · [来源 2](<https://arxiv.org/abs/2606.20493>) | A | 沿用既有阅读记录 |
| 544 | [Liquid AI LFM2.5 Retrievers：本地多语言 memory/search 检索底座](<https://www.liquid.ai/blog/lfm2-5-retrievers>) | A | 沿用既有阅读记录 |
| 545 | [xAI Grok Build /goal：long-running coding agent 的目标状态与验证面](<https://x.ai/news/introducing-goal>) | A | 沿用既有阅读记录 |
| 546 | [Claude Tag：Slack 中的 scoped team agent 与组织级权限/成本控制](<https://www.anthropic.com/news/introducing-claude-tag>) | A | 沿用既有阅读记录 |
| 547 | [The Coming Loop：harness-level loop 会放大工程质量债](<https://lucumr.pocoo.org/2026/6/23/the-coming-loop/>) | A | 沿用既有阅读记录 |
| 548 | [Mistral OCR 4 + Baidu Unlimited-OCR：文档 ingestion 从 OCR 走向结构化 context substrate](<https://mistral.ai/news/ocr-4/>) · [来源 2](<https://github.com/baidu/Unlimited-OCR>) | A | 沿用既有阅读记录 |
| 549 | [Randomized YaRN：短上下文训练也能改善 16K-128K 长上下文推理泛化](<https://arxiv.org/abs/2606.23687>) | A | 沿用既有阅读记录 |
| 550 | [AIR：用 RL 学会何时在多模态推理中调用代码工具](<https://arxiv.org/abs/2606.23678>) | A | 沿用既有阅读记录 |
| 551 | [Can LLMs Reliably Self-Report Adversarial Prefills：模型自我报告不能当安全证据](<https://arxiv.org/abs/2606.23671>) | A | 沿用既有阅读记录 |
| 552 | [VibeThinker-3B：小模型 verifiable reasoning 的成本/能力边界](<https://arxiv.org/abs/2606.16140>) | A | 沿用既有阅读记录 |
| 553 | [Gemini 3.5 Flash computer use：UI agent 的 observe-act-screenshot contract](<https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-computer-use-gemini-3-5-flash/>) | A | 沿用既有阅读记录 |
| 554 | [Qwen-AgentWorld：language world model for agentic RL](<https://arxiv.org/abs/2606.24597>) · [来源 2](<https://github.com/QwenLM/Qwen-AgentWorld>) | A | 沿用既有阅读记录 |
| 555 | [AOHP：Android Open Harness Project / OS-level agent harness](<https://arxiv.org/abs/2606.23449>) · [来源 2](<https://github.com/aohp-os/aohp>) | A | 沿用既有阅读记录 |
| 556 | [OpenThoughts-Agent：agentic model data recipes](<https://arxiv.org/abs/2606.24855>) · [来源 2](<https://github.com/open-thoughts/OpenThoughts-Agent>) | A | 沿用既有阅读记录 |
| 557 | [Tmax：terminal-agent RL recipe and TMax-15K](<https://arxiv.org/abs/2606.23321>) · [来源 2](<https://github.com/hamishivi/tmax>) · [来源 3](<https://huggingface.co/allenai/tmax-9b>) | A | 沿用既有阅读记录 |
| 558 | [OpenAI/Broadcom Jalapeno：inference cost as agent runtime constraint](<https://openai.com/index/openai-broadcom-jalapeno-inference-chip/>) | A | 沿用既有阅读记录 |
| 559 | [Qualcomm 收购 Modular：portable AI serving/software stack signal](<https://investor.qualcomm.com/news-events/press-releases/news-details/2026/Qualcomm-to-Acquire-Modular/default.aspx>) | A | 沿用既有阅读记录 |
| 560 | [Headroom：agent context compression / CCR gate](<https://github.com/headroomlabs-ai/headroom>) | A | 沿用既有阅读记录 |
| 561 | [Notion Mail agent takeover + General Intuition action-labeled world model data](<https://techcrunch.com/2026/06/25/notion-mail-shuts-down-amid-agent-takeover/>) | A | 沿用既有阅读记录 |
| 562 | [OpenAI GPT-5.6 Sol limited preview：frontier model release gate and ultra subagents](<https://openai.com/index/previewing-gpt-5-6-sol/>) | A | 沿用既有阅读记录 |
| 563 | [OpenAI Codex economic research：agents transform work into delegated long-horizon tasks](<https://openai.com/index/how-agents-are-transforming-work/>) | A | 沿用既有阅读记录 |
| 564 | [Cursor reward hacking in coding benchmarks：strict harness for aware coding agents](<https://cursor.com/blog/reward-hacking-coding-benchmarks>) | A | 沿用既有阅读记录 |
| 565 | [Workweave Router：cache-aware model routing for Claude Code / Codex / Cursor](<https://github.com/workweave/router>) | A | 沿用既有阅读记录 |
| 566 | [NVIDIA NeMo AutoModel / Transformers v5 Expert Parallelism + DeepEP MoE fine-tuning path](<https://huggingface.co/blog/nvidia/accelerating-fine-tuning-nvidia-nemo-automodel>) · [来源 2](<https://github.com/deepseek-ai/DeepEP>) | A | 沿用既有阅读记录 |
| 567 | [PEEU GUI agents：Autonomous Experience Exploration + Hindsight Experience Utilization for task planning](<https://arxiv.org/abs/2606.27330>) | A | 沿用既有阅读记录 |
| 568 | [When are likely answers right? Sequence Probability and Correctness in LLMs](<https://arxiv.org/abs/2606.27359>) | A | 沿用既有阅读记录 |
| 569 | [Un-0：open coupled-oscillator image generator as physical-compute substrate](<https://unconv.ai/blog/introducing-un-0-generating-images-with-coupled-oscillators/>) · [来源 2](<https://github.com/unconv-ai/Un-0>) | A | 沿用既有阅读记录 |
| 570 | [OpenAI + Broadcom Jalapeno inference chip：full-stack LLM inference platform](<https://openai.com/index/openai-broadcom-jalapeno-inference-chip/>) | A | 沿用既有阅读记录 |
| 571 | [DeepSeek DSpark / DeepSpec：confidence-scheduled speculative decoding for production serving](<https://github.com/deepseek-ai/DeepSpec>) · [来源 2](<https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-DSpark>) | A | 沿用既有阅读记录 |
| 572 | [AWS Lambda MicroVMs：full-lifecycle Firecracker sandboxes for AI agents](<https://aws.amazon.com/blogs/aws/run-isolated-sandboxes-with-full-lifecycle-control-aws-lambda-introduces-microvms/>) | A | 沿用既有阅读记录 |
| 573 | [DBOSify：Postgres-backed durable workflow as compact Temporal alternative](<https://github.com/dbos-inc/dbosify-py>) | A | 沿用既有阅读记录 |
| 574 | [Adrafinil：macOS activity assertion layer for long-running AI agents](<https://github.com/kageroumado/adrafinil>) | A | 沿用既有阅读记录 |
| 575 | [Cloud World Model：cloud-infra simulation product radar](<https://www.producthunt.com/>) | A | 仅元信息；原记录仅有聚合入口；未补齐具体原文，不作为已读材料。 |
| 576 | [Are We Ready For An Agent-Native Memory System? Trustworthy Memory Search](<https://arxiv.org/abs/2606.24775>) · [来源 2](<https://github.com/HKUDS/MemorySearch>) | A | 沿用既有阅读记录 |
| 577 | [Semgrep GLM-5.2 cyber benchmark：real security harness for coding agents](<https://semgrep.dev/blog/2026/we-have-mythos-at-home-glm-52-beats-claude-in-our-cyber-benchmarks>) | A | 沿用既有阅读记录 |
| 578 | [GitHub Copilot App BYOK：provider/account boundary for agent sessions](<https://github.blog/changelog/2026-06-23-github-copilot-app-support-for-byok/>) | A | 沿用既有阅读记录 |
| 579 | [Wayfinder Router：deterministic local/cloud LLM routing](<https://github.com/itsthelore/wayfinder-router>) | A | 沿用既有阅读记录 |
| 580 | [OpenAI Codex issue：exclude sensitive files by default](<https://github.com/openai/codex/issues/2847>) | A | 沿用既有阅读记录 |
| 581 | [Tokenmaxxing / agent output budget policy](<https://news.ycombinator.com/item?id=48699302>) | A | 沿用既有阅读记录 |
| 582 | [百度千帆 Coding Plan -> Token Plan：coding agent usage ledger signal](<https://cloud.baidu.com/article/3730945>) | B | 沿用既有阅读记录 |
| 583 | [Dual Strix Halo vLLM cluster：local serving lab reference](<https://github.com/recallnet/strix-halo-vllm>) | B | 沿用既有阅读记录 |
| 584 | [Claude Sonnet 5：agent default model cost-performance reset](<https://www.anthropic.com/news/claude-sonnet-5>) | A | 沿用既有阅读记录 |
| 585 | [Claude Science AI workbench：auditable domain agent OS](<https://www.anthropic.com/news/claude-science-ai-workbench>) | A | 沿用既有阅读记录 |
| 586 | [Gemini Spark updates：desktop automation + connected apps + custom MCP](<https://blog.google/innovation-and-ai/products/gemini-app/gemini-spark-updates-june-2026/>) | A | 沿用既有阅读记录 |
| 587 | [vLLM Micro-Agent：serving router as bounded agent collaboration](<https://vllm.ai/blog/2026-06-29-micro-agent-frontier-models>) | A | 沿用既有阅读记录 |
| 588 | [SWE-Together：interactive user-session benchmark for coding agents](<https://arxiv.org/abs/2606.29957>) | A | 沿用既有阅读记录 |
| 589 | [OSWorld 2.0：long-horizon computer-use official runner boundary](<https://arxiv.org/abs/2606.29537>) | A | 沿用既有阅读记录 |
| 590 | [SWE-MeM：adaptive memory management for long-horizon coding agents](<https://arxiv.org/abs/2606.28434>) | A | 沿用既有阅读记录 |
| 591 | [Language Firewall：routing defense for multi-agent systems](<https://arxiv.org/abs/2606.30555>) | A | 沿用既有阅读记录 |
| 592 | [Couchbase AI Data Plane：enterprise memory/context substrate](<https://www.couchbase.com/blog/your-ai-agents-are-stuck-in-pilot-its-a-data-problem-not-a-model-problem/>) | A | 沿用既有阅读记录 |
| 593 | [Lingtai：local-first lifelong Agent runtime](<https://github.com/Lingtai-AI/lingtai>) | A | 沿用既有阅读记录 |
| 594 | [Ephemeral Sandbox：COW workspace 与 OCC publication](<https://github.com/Ephemeral-AI-Lab/ephemeral-sandbox>) | A | 沿用既有阅读记录 |
| 595 | [Using Claude Code: The Unreasonable Effectiveness of HTML](<https://x.com/trq212/status/2052809885763747935>) | A | 沿用既有阅读记录 |
| 596 | [梁文锋投资者交流会 · 网传录音文字稿（非官方）](<https://drive.google.com/file/d/13NSXtglkcMJoWPRyT_2yQq_VhREG8nkt/view>) · [来源 2](<https://www.nbd.com.cn/articles/2026-07-23/4504599.html>) | A | 仅元信息；已取得 42 页 PDF、提取文本并检查首页，未逐段精读。网传转写稿，未经 DeepSeek 或本人公开确认；ASR 与 AI 整理可能引入错误，不作为官方口径。 |
| 597 | [hai-stack / Geju：用 target model、falsifier 与收益账单抵抗局部补丁化](<https://github.com/hylarucoder/hai-stack>) | A | 沿用既有阅读记录 |
| 598 | [Waza：把工程习惯产品化为可路由、可验证的 Agent skills](<https://github.com/tw93/Waza>) | A | 沿用既有阅读记录 |
| 599 | [BfdCampos Mermaid skill：把“语法正确”提升为“渲染后可读”](<https://github.com/BfdCampos/dotfiles>) | A | 沿用既有阅读记录 |
| 600 | [Oracle：把第二模型意见封装成可追踪的 consult session](<https://github.com/steipete/oracle>) | A | 沿用既有阅读记录 |
| 601 | [Graph Engineering：给 Agent Harness 叠加显式控制流图，而不是换一个新名词](<https://x.com/0xCodez/status/2081429287945506950>) | A | 沿用既有阅读记录 |
| 602 | [赵克常《炒股挣钱》：风险认知课，不是可复制的投资研究方法](<https://x.com/ActionXAi/status/2083691580196856018>) | B | 沿用既有阅读记录 |
| 603 | [未核实的 MLSys 截图线索：集中式推理、确定性信号与知识图谱控制面](<https://mlsys.org/Conferences/2026/CallForResearchPapers>) | B | 沿用既有阅读记录；来源真实性未确认；保留为核验案例，不作为论文结论。 |
| 604 | [Evolvent AI GitHub 组织复核：与既有研究目录的重复记录](<https://github.com/evolvent-ai>) | B | 沿用既有阅读记录；保留稳定 ID 与重复记录；不因重复而静默删除。 |