Quickstart

A five-minute happy path: launch a fleet, let it learn your codebase, ask it to do something, watch the workers, and exit. This assumes you've already installed agents-fleet and set up a provider.

1. Launch the REPL

From inside any project directory:

cd your-project
agents-fleet

You'll land in an interactive REPL with a status bar showing the active model and worker cap:

❯ Type a prompt or /command... (/ + Tab for commands)
  model: claude-opus-4.6   workers: 12                    ✦ agents-fleet

Discover commands

Type / then press Tab to open the interactive command picker with fuzzy search. There are ~40 slash commands — see the Commands Manual.

2. Profile the repo with /init

Before asking the fleet to do real work, let it understand your codebase:

❯ /init

This spawns seven parallel explorers plus one adversarial critic. Each explorer investigates a different domain — architecture, build system, patterns, decisions, dependencies, APIs, security — and the critic challenges every finding. Results are written to .fleet/context/ and auto-loaded in every future session, so you only run /init once per project (re-run with /init refresh after big changes).

You'll see live progress as the explorers work, then a synthesized summary of your project's architecture and conventions.

3. Ask the coordinator to do something

Just type a request in plain English. The coordinator decides how to break it down and spawns workers as needed:

❯ Add input validation to the user-registration endpoint and a test for it

The coordinator plans the work, spawns one or more workers (for example a coder to make the change and a tester to add coverage), tracks their dependencies, and reports back when they finish. Coder workers run in isolated git worktrees, so parallel changes never collide.

For a heavier, planned task, use a command instead of free-form prompting:

❯ /fast-do "fix the failing test in auth.test.ts"     # quick — skips planning, up to 3 workers
❯ /do "refactor the auth module to use dependency injection"   # full plan → execute → verify

4. Watch the fleet

While workers run, a chip strip appears along the bottom of the REPL showing the coordinator plus every active worker. Check progress without leaving the prompt:

❯ /status     # one-shot snapshot: workers, tasks, tokens
❯ /workers    # detailed per-worker state
❯ /tasks      # the task DAG and what's blocked/ready

Switch into a worker's transcript with Ctrl+] / Ctrl+[ or /view <name>. Worker views are read-only in the current release — you observe, but steer through the coordinator. See Worker Views for the full keybinding reference.

5. Exit

When you're done:

❯ /exit

/exit (also /quit, /q) runs a graceful shutdown — it extracts steering insights, synthesizes learnings, and saves the session so you can pick up later with agents-fleet --resume.

What just happened?

You launched a coordinator that profiled your repo, then dispatched short-lived workers to do focused work, tracked as a task DAG. That's the core loop. To understand the pieces, read Core Concepts; for a guided tour of what's worth trying next, see Your First Fleet.