Composio Agent Orchestrator (AO) is an open-source system that manages fleets of parallel AI coding agents — each working in its own git worktree — and autonomously routes CI failures, merge conflicts, and code review comments back to the responsible agent, without human intervention.
The Problem It Solves
Running a single AI coding agent is table stakes. The hard part is running ten of them simultaneously on ten different issues, without conflicts, without manual babysitting, and with a clear escalation path when something goes wrong. AO treats that coordination problem as a first-class system concern.
Its design philosophy is push-not-pull: the human never polls a dashboard to see what's happening. Instead, notifications arrive when human judgment is genuinely required (PR ready to merge, agent stuck after retries). Everything else — CI failures, test flaps, review change requests — is handled automatically.
Architecture in Brief
AO is built around two key ideas:
Session state machine: Every agent session progresses through well-defined states (
SPAWNING → WORKING → PR_OPEN → CI_FAILED | REVIEW_PENDING → MERGEABLE → MERGED). The Lifecycle Manager drives transitions and fires reactions — so aCI_FAILEDevent automatically triggers a retry agent, not a Slack ping to a developer.Eight plugin slots: Each architectural concern is a swappable plugin — runtime (tmux, Docker, Kubernetes, E2B), agent (Claude Code, Codex, Aider, Goose), workspace (git worktree, clone), tracker (GitHub, Linear, Jira, GitLab), SCM, notifier, terminal, and lifecycle. Default implementations get you running with one command; plugins make it production-ready.
ao start https://github.com/your-org/your-repo
Two-Tier Notification Model
AO minimizes interruptions with a two-tier escalation model:
- Auto-handled: CI failures, review comments, merge conflicts — AO retries automatically and escalates only after repeated failures
- Always notify: Agent stuck, PR ready to merge, crash, auto-fix failure — these require human judgment and always get pushed immediately
Why Trial
AO has earned its place in trial by solving a real coordination problem with a clean, extensible design:
- 5.7k GitHub stars within months of open-sourcing (February 2026), with 752 forks and 622 commits
- 3,288 test cases and an active release cadence (v0.2.2 as of March 2026)
- Agent-agnostic: Supports Claude Code, Codex, Aider, Goose — your choice
- Tracker-agnostic: Works with GitHub, GitLab, Linear, Jira — brings AO into existing workflows
- MIT license with no vendor lock-in
The caveat: AO is a young project (sub-1.0). The plugin system is well-designed but some slots (Docker, Kubernetes runtimes, enterprise auth) are roadmap items, not yet shipped. Teams needing production-grade fleet orchestration today should audit the Phase 2/3 roadmap against their requirements.
When AO Makes Sense
- You're running background coding agents across a large issue backlog
- You want autonomous CI remediation without writing custom glue code
- You need an abstraction layer that works with multiple agent tools
- You want to minimize context switching by receiving only actionable notifications
Not Right For
- Building custom agent logic (use CrewAI, LangGraph, or Google ADK instead)
- Single-agent workflows where the overhead isn't justified
- Teams that need enterprise auth, audit logging, or SSO today (Phase 4 roadmap)
Key Characteristics
| Property | Value |
|---|---|
| Language | TypeScript |
| License | MIT |
| GitHub | ComposioHQ/agent-orchestrator |
| Stars | ~5.7k (April 2026) |
| Provider | Composio |
| Version | 0.2.2 |
| First Release | February 2026 |
See Also
- Architecture Deep Dive
- Composio — the tool integration platform from the same team
- Background Coding Agents — the broader practice this tool operationalizes