Introduction
A multi-agent AI swarm in your terminal. Point it at any codebase, and it orchestrates teams of specialized AI workers to investigate, plan, build, test, and review code — all while you watch.
What is agents-fleet?
agents-fleet is a CLI tool that turns a single AI session into a coordinated team of up to 12 parallel workers. Instead of one assistant doing everything sequentially, a long-lived coordinator agent breaks work into tasks, spawns specialized workers, tracks their dependencies as a graph, and synthesizes the results back into one coherent answer.
Think of it as a software-engineering team that:
- Investigates your codebase with seven parallel explorers plus an adversarial critic (
/init). - Plans features through a multi-phase pipeline — brainstorm → requirements → research → design → validation → definition-of-done → task breakdown (
/feature). - Builds code in isolated git worktrees, so workers never trample each other's changes.
- Reviews its own output with four simultaneous reviewers — security, correctness, architecture, and performance (
/code-review). - Runs unattended via autopilots and goal-driven loops that keep working toward an objective and open PRs when done.
- Learns from every session, evolving its skill prompts over time from real telemetry.
Dual-provider support
You pick the underlying LLM at launch:
- GitHub Copilot (default) — uses your GitHub Copilot subscription through the
ghCLI. Launch withagents-fleetoragents-fleet --provider copilot. - Anthropic Claude — uses the Claude Agent SDK via your Anthropic credentials. Launch with
agents-fleet --provider claude.
Both providers expose the same orchestration surface. Sessions are pinned to the provider they were created with, so resuming a Claude session under Copilot (or vice-versa) is rejected with a clear error. See Installation for auth setup.
Source of truth
Provider selection is defined in src/index.ts (--provider copilot|claude). Worker concurrency defaults to 12 (--max-workers).
The extension surface
agents-fleet is composable. Beyond the built-in commands, you can author your own building blocks across three override tiers (bundled → user → project):
| Artifact | What it is |
|---|---|
| Roles | A coordinator or worker persona with a declared agent type; runtime policy determines available tools and permissions. |
| Skills | Atomic, reusable units of knowledge composed into a worker's prompt. |
| Workflows | Multi-stage execution plans with artifacts and parallel/sequential stages. |
| Crews | Named teams that bind roles, skills, and a workflows: override map. |
Bundled and user artifacts load automatically. Project roles, skills, and crews are default-off and require --trust-project-skills; project workflows use a separate trust control. The Composition Guide is the canonical reference for prompt order, runtime-owned invariants, trust, and reload behavior.
External tools plug in through the Model Context Protocol (MCP), and a control plane lets you observe and drive long-running fleets remotely (for example, over Telegram).
Where to go next
- Installation — get it running.
- Quickstart — your first five minutes.
- Core Concepts — the mental model behind the swarm.
Built on the GitHub Copilot SDK and the Anthropic Claude Agent SDK.