---
title: "How Do Agents Fail on AutoResearch: End-to-End Diagnostic Evaluation on 100 Real-World Frontier Research Tasks"
type: paper-summary
source_paper: "2026-08-14_21-39-38Z_HowDoAgentsFailonAutoResearch_End_to_EndDiagnostic.md"
---
# Summary: How Do Agents Fail on AutoResearch

**Original paper:** [arXiv: 2608.14905](http://arxiv.org/abs/2608.14905v1)

## Summary
AutoResearchEval evaluates 100 frontier research tasks, 800 agent trajectories, and eight harness-model combinations across seven domains. Its 45-pattern failure taxonomy finds a common metacognitive deficit: agents often cannot verify outputs against evidence or revise after detecting errors.

## Why it matters
Autonomous research needs process-level evaluation and artifact-grounded verification, not just a final-answer score. The paper identifies the missing control loop directly.