Herding cats, but they’re AI agents that fix their own CI

What it does Agent Orchestrator manages fleets of AI coding agents working in parallel on a single codebase. Each agent gets its own git worktree, branch, and pull request, while the system autonomously handles CI failures, review comments, and merge conflicts. You review outcomes from a local dashboard and intervene only when human judgment is required.

The interesting bit The project treats agent coordination as an infrastructure problem rather than a prompt-engineering hack. It exposes a seven-slot plugin architecture—runtime, agent, workspace, tracker, SCM, notifier, terminal—that lets you swap Claude Code for Codex, GitHub for Linear, or tmux for Docker without rewriting the core orchestration logic.

Key highlights

• Every agent is isolated in its own git worktree, so experiments and branches never collide.
• Feedback loops like ci-failed or changes-requested are automatically routed back to the originating agent with configurable retries and escalation timers.
• The orchestrator ships with a web dashboard and a CLI (ao) that auto-generates a validated YAML configuration file.
• On macOS, it holds a caffeinate assertion to keep the machine awake for remote dashboard access; Linux and Windows lack this feature for now.
• Seven plugin interfaces allow swapping agents, runtimes, issue trackers, and notification channels.

Caveats

• macOS lid-close sleep is enforced at the hardware level and cannot be overridden; remote access with the lid shut requires clamshell mode.
• Idle-sleep prevention is currently unavailable on Linux and Windows, leaving remote access vulnerable to OS power management.

Verdict This is for teams already using Claude Code or Codex who are drowning in branch management and want agents to handle their own CI tantrums. If you are uncomfortable with autonomous agents pushing commits and addressing reviews while you sleep, pass.