Introduction

Agent Orchestrator spawns parallel AI coding agents — one per issue, each in its own worktree with its own PR — and gives you one dashboard to supervise them all.

Agent Orchestrator (AO) runs many AI coding agents at once. Point it at a repo, hand it an issue, and it spawns an agent in an isolated git worktree. You watch the work on a Kanban-style dashboard, merge when ready.

Every interesting piece is a plugin — the agent (Claude Code, Codex, Cursor, Aider, OpenCode), the runtime (tmux or a plain child process), the tracker (GitHub, GitLab, Linear), the notifier (Slack, Discord, desktop, webhook). You pick what you want in agent-orchestrator.yaml and AO wires it together.

AO can also run different agents per role — a reasoning-heavy model for the orchestrator that supervises sessions, and a fast one for the workers that actually write code. See Per-role agents.

Fastest path

Install with one command, then spawn your first agent in under five minutes.

What it does for you

Instead of…	AO does this
Running one agent at a time, waiting for it to finish	Runs N agents in parallel, one per issue, each isolated
Babysitting CI and manually re-running agents when red	Detects CI failures and wakes the agent to fix them
Copy-pasting review comments back to the agent	Spots "changes requested" and nudges the agent automatically
Losing your place every time you close your laptop	Persists everything under `~/.agent-orchestrator`, resumable across reboots

How it fits together

issue  ─►  AO spawn  ─►  worktree + agent  ─►  PR  ─►  CI / review loop  ─►  merged
                              │                                  ▲
                              └──────── dashboard (SSE) ─────────┘

Each moving part is swappable:

Agent — which AI tool actually writes code
Runtime — where it runs (tmux window on macOS/Linux, child process on Windows)
Workspace — worktree or full clone
Tracker — where the issues and PRs live
Notifier — who gets pinged when the agent needs you

See Plugins for the full catalog.