1. One agent, one turn
Start with one Linux machine. It runs the fleet server. You give the server a task. The server runs Claude Code on it and keeps what happens.
- An agent is one named task. It has a brief, a project, a working
directory and an owner.
fleet spawn <name>makes it. The brief comes on stdin. - A turn is one run of
claude -pfor one agent. Turn 1 gets the brief and a context block that tells the agent how to report. A later turn continues the same conversation with--resume(fleet/internal/driver/claude/claude.go). - Inside the turn, the agent reports with
fleet statusand finishes withfleet done. It can use thefleetcommand or the same tools in the worker MCP. - The server keeps every agent in the registry, every change in
the ledger, and each brief and conversation as a
blob. All three are in
/var/lib/fleeton the server. - You see the result with
fleet ls, or on the page.
One machine
On one machine, the server starts each turn itself. A turn runs under fleet turn-shim. If the
server restarts, the turn continues. The scheduler looks for work
every 15 s, and also when something wakes it.
Install: Quickstart.
Next: 2. The life of an agent. A turn does not always end with done.