ComposioHQ Agent Orchestrator

github.com/ComposioHQ/agent-orchestrator

What it is

Spawn parallel AI coding agents, each in its own git worktree. Agents autonomously fix CI failures, address review comments, open PRs. Agent-agnostic (Claude Code, Codex, Aider). Runtime-agnostic (tmux, Docker). Tracker-agnostic (GitHub, Linear).

What we kept

• Git worktree per agent. Verbatim. This is the source of our branchIsolation.useWorktrees config and the worktree_* shell functions in run.sh.
• Branch-per-feature with PR or merge on pass. Our onPass: merge|pr|keep mirrors their flow.
• Auto-PR creation on pass. We use gh pr create to wire this up.

What we dropped

• Agent-agnostic dispatch. We’re Claude-only. Adding Codex/Aider support would require abstracting the invoke interface; we don’t need it.
• CI fix automation as a first-class concern. Our analogue is the validator + debugger flow — failed validations trigger debugger investigations. CI per se is out of scope (no shared CI in our missions).
• Tracker integration (GitHub/Linear). User explicitly rejected external task sources. Issues live in .harness/issues.json.

Differences in philosophy

Composio	Papercup
Agent-agnostic dispatch	Claude-only
CI/PR/review-comment automation	Internal validator + worktree merge
Tracker-driven (GitHub/Linear)	File-driven (features.json)
tmux or Docker runtime	Bash subprocess

Honest take

Composio is the most production-pragmatic framework we surveyed. If you’re a team that already has a GitHub-based workflow and wants AI to do the merge-conflict and CI-fix dance, Composio is what you want.

We took the worktree pattern (it’s correct) and skipped everything else. The tracker integrations are worth their weight if you have an existing tracker; we didn’t.