Managing Agents

Running one agent is a command. Running many is a fleet — and a fleet needs somewhere to watch it from.

There are three entry points, and they are views onto the same running machine.

The CLI

Start an agent, run a prompt, check what's alive.

The TUI

Talk to an agent. Switch between them. One terminal, whichever agent you want.

The extension

Axon Fleet puts the whole machine in the editor: every agent, every running process, every past run, and the console to inspect any of them.

They share one runtime

Every booted agent registers itself locally. All three read that same registry, so none of them owns a fleet — they observe one.

Start an agent from a terminal and it appears in the sidebar. Click it in the sidebar and you can take over the conversation in the TUI, in the session it was already in.

That is the integration: you are never restarting an agent to look at it from somewhere else. You are joining a run in progress.

Two views

The extension organises the machine into two.

Assets — what can do work. Jobs — what needs doing.

The whole interface exists to make the edge between them cheap: take something from the register, point it at something on the board.

The human assigns

Agent capacity is effectively free, so scheduling agents efficiently buys nothing. The scarce resources are your review bandwidth and your codebase's integrity — so a person decides what runs against what.

Fleet reads your machine, not your account. Agents come from disk, processes from the local registry, jobs from your repos. Nothing is uploaded.

Next: Running Instances — everything alive, and what you can do to it.