# Roadmap

This is what's next for AgentsCommander. Items are grouped by status: **Shipped** / **Planned** / **Considering**. Order within each section is rough priority.

This file is a snapshot. The authoritative status for tracked work lives in its linked GitHub issue; current installation support lives in the [platform contract](docs/install-with-agent.md#support-gates).

## Shipped (highlights)

- Multi-agent rooms with file-based inter-agent messaging
- Cross-coding-agent profiles: Claude Code · Codex · Pi · Antigravity
- OpenCode usable today via the custom coding-agent path; configure it with any provider or model OpenCode supports
- Agents Agency role-template picker (explicit downloaded cache from [@msitarzewski/agency-agents](https://github.com/msitarzewski/agency-agents))
- Telegram bridge (PTY mode for every session, provider-specific JSONL readers where available, plus photo/image sends for screenshots and artifacts)
- Voice-to-text via Gemini API with auto-execute
- Portable instances (rename the `.exe` → isolated workspace, ports, mutex, config)
- Embedded HTTP/WebSocket server (per-instance, opt-in)
- Step-by-step PTY visibility per agent (xterm.js + WebGL + ConPTY)

## Planned

### Coding-agent integrations

- **OpenCode first-class profile** - OpenCode already runs via the custom coding-agent path (provider-agnostic). Remaining work: a tuned `CodingAgentKind` profile with resume tokens and idle tuning, alongside Claude Code, Codex, Antigravity, and Pi. ([#315](https://github.com/mblua/AgentsCommander/issues/315))
- **NVIDIA OpenShell** - sandbox-runtime integration (not a coding-agent profile; OpenShell hosts other coding agents inside a policy-driven sandbox). ([#316](https://github.com/mblua/AgentsCommander/issues/316))

### Per-agent configuration

- **Per-agent model + effort-level selection** - pin a specific model (e.g. Opus 4.7 vs Sonnet 4.6) and a thinking-effort tier (low / medium / high / ultra) to each agent independently. Today the choice is per-session at launch time; this would persist per `_agent_<name>` in the matrix and survive across rooms. ([#329](https://github.com/mblua/AgentsCommander/issues/329))

### Organigram & scaling

- **Build the company - multi-level orchestrator hierarchies** - A room today is one layer. The next step layers them: level-A orchestrator talks to its level-B reports; B talks to C; A never reaches C directly. Skip-levels flow through the chain, like in any sane org. The result: a real organigram, enforced by the messaging topology instead of by convention. Cascade decisions down, roll status up, hold accountability at 5, 50, or 500 agents the same way. ([#330](https://github.com/mblua/AgentsCommander/issues/330))
- **Curated role-team library** - ship a base catalog of task-ready teams, not just isolated roles. Each template defines the orchestrator, specialist roles, task decomposition pattern, role boundaries, quality gates, and handoff contracts so a user can delegate a goal to a pre-curated team that can execute end to end with predictable accountability.
- **Orchestrator auto-handoff for context management** - when an orchestrator reaches a configurable context threshold, it can write a compact handoff of active continuity state and continue fresh. Forgettable or exclude-from-handoff details stay out. ([#349](https://github.com/mblua/AgentsCommander/issues/349))
- **Telegram conversation routing between orchestrators** - transfer a Telegram conversation from one orchestrator/room to another without manual bot or channel reconfiguration. ([#350](https://github.com/mblua/AgentsCommander/issues/350))

### CLI capabilities

- **`create-room` / `create-team` / `create-project`** - non-interactive equivalents of the existing `create-agent` flow for headless / script-driven setup. ([#317](https://github.com/mblua/AgentsCommander/issues/317))

### Automation

- **Cron-based scheduled executions** - define explicit, inspectable, and auditable cron schedules that regularly trigger events for an orchestrator, room, agent, or workflow target, supporting recurring status checks, reports, maintenance tasks, and scheduled room runs. ([#354](https://github.com/mblua/AgentsCommander/issues/354))

### Execution determinism

- **AC Harness** - deterministic command-execution layer for coding-agent tool calls, extending beyond Claude to Codex, and future agents (today only Claude's `.claude/settings.local.json` hook is wired). ([#318](https://github.com/mblua/AgentsCommander/issues/318))

### Interoperability

- **Export AC agent configs** to **LangChain Deep Agents**, **OpenAI Agents SDK**, and **CrewAI** - author an agent once in AC, run it programmatically anywhere. ([#319](https://github.com/mblua/AgentsCommander/issues/319))

### Platform

- **Linux support expansion** - Linux x86_64/AMD64 is supported with medium test coverage. Broaden distribution and feature coverage, including gaps where capabilities remain Windows-only. Testers can use the [reproducible platform report](docs/install-with-agent.md#help-extend-linux-and-macos-support).
- **macOS support and verification** - macOS is supported via npm only, with very low test coverage, and errors are expected. Grow the coverage and fix what reproducible reports surface. Testers and contributors are welcome through the [collaboration path](docs/install-with-agent.md#help-extend-linux-and-macos-support); artifact availability does not change the support tier.
- **Web landing page** on `agentscommander.org` or GitHub Pages from `/docs`. ([#321](https://github.com/mblua/AgentsCommander/issues/321))

## Considering

- **Discord community** - defer until star count justifies it (~500 stars). ([#322](https://github.com/mblua/AgentsCommander/issues/322))
- `docs/recipes/` - end-to-end multi-agent walkthroughs
- **`cargo udeps` cleanup** - audit and trim unused crates (start with `futures` vs `futures-util`). ([#323](https://github.com/mblua/AgentsCommander/issues/323))

---

Have an idea? Open a GitHub Discussion under **Ideas** or comment on the relevant issue above.