Overstory is an open-source multi-agent coding orchestration system that runs a coordinated team of specialized agents (Scout, Builder, Reviewer, Merger) in isolated Git worktrees via tmux, coordinated through a typed SQLite mail system. It's the most runtime-agnostic team-model orchestrator in this space, with 11 pluggable adapter backends.
Why It's in Assess
Overstory occupies the same design space as Taskplane — a genuine team model rather than a fleet model. Multiple agents with specialized roles collaborate on the same codebase: Scout explores read-only, Builder implements, Reviewer validates, Merger integrates. The key differentiator over Taskplane is runtime portability: a standardized AgentRuntime interface lets you swap between Claude Code, Pi, Codex, Gemini, Aider, Cursor, Goose, Amp, and more without changing orchestration logic.
The project documents its risks honestly in STEELMAN.md, explicitly warning that "agent swarms are not a universal solution" and enumerating compounding error rates, cost amplification, and merge conflict prevalence as expected failure modes — not edge cases. That transparency is a positive signal for production-mindedness.
In Assess (not Trial) because: only Claude Code and Sapling adapters are stable — the other 9 are experimental; setup requires tmux + Bun + SQLite; the documentation advises reading a companion book before production deployment; and 1.2k stars is meaningful but not yet the broader adoption signal that would justify Trial.
Architecture
Agent hierarchy:
Orchestrator (multi-repo) → Coordinator (project-level) → Lead → Scout / Builder / Reviewer / Merger
A three-tier Watchdog (mechanical daemon → AI triage → Monitor agent) handles health monitoring and crash recovery. The Supervisor role is listed as deprecated in favour of the Lead/Coordinator split.
Inter-agent communication: Custom SQLite mail system in WAL mode (~1–5ms queries) with 8 typed message types (worker_done, merge_ready, dispatch, escalation, etc.) and group addressing (@all, @builders). Agents communicate asynchronously without shared filesystem state.
Tool enforcement: Runtime-specific guards mechanically block non-implementation agents from file writes and dangerous git operations. Claude Code uses settings.local.json hooks; Pi uses .pi/extensions/ guard extensions. Guard coverage varies by runtime — several experimental adapters lack enforcement.
Merge queue: FIFO queue with 4-tier conflict resolution merges agent branches sequentially into canonical.
Key Features
- 11 runtime adapters: Claude Code (stable), Sapling (stable), Pi/Copilot/Cursor/Codex/Gemini/Aider/Goose/Amp/OpenCode (experimental)
- Instruction overlays: Base
.mdworkflow definitions inject into per-task agent overlays — separation of process from task scope - Checkpoint/handoff: Session save/restore for context compaction survivability and crash recovery
- os-eco ecosystem: Coordinates with Mulch (codebase analysis) and Canopy (prompt rendering)
- Gateway providers: Route API calls through OpenRouter, z.ai, or self-hosted proxies
Caveats
- Most adapters (9 of 11) are experimental and lack mechanical tool enforcement
- Requires tmux and Bun (not standard Node.js) — non-trivial setup
- Multi-agent debugging complexity is significantly higher than single-agent workflows — acknowledged explicitly by the project
- Solo author (jayminwest); no enterprise backing
- Production use requires understanding documented failure modes (STEELMAN.md + companion book)
Key Characteristics
| Property | Value |
|---|---|
| Interface | CLI + ANSI TUI dashboard |
| License | MIT |
| Underlying model | Any (11 runtime adapters) |
| Isolation | Git worktrees (per agent, via tmux) |
| Coordination | SQLite mail system |
| Pricing | Free (BYOK) |
| Provider | jayminwest (individual) |
| GitHub | jayminwest/overstory |
| npm | @os-eco/overstory-cli |