Keep a watcher on your agents

A watcher is an agent that checks on your other agents for you, so you don’t have to look in on each one. You give it a roster, the list of agents it watches, and it looks at each one’s terminal on a schedule.

A watcher is an agent, not a script. It can reason about what it sees, and you change what it does by sending it a message in plain words.

Start a watcher

Your coordinator is the agent you talk to. It starts the other agents, hands out work and reports back to you. Thrum’s web UI is the page you open in your browser to see your agents and their messages. My Inbox is your inbox in Thrum’s web UI. Ask your coordinator, in its terminal or in My Inbox, and name the agents to watch:

Start a watcher on my implementers.

The coordinator starts the watcher and sets how often it checks. It may ask which model to run the watcher on.

Tell it what to change

To change what the watcher does, send it a message. In My Inbox, click + Add, pick the watcher and write to it.

For example, when a new agent starts, ask the watcher to add it to its roster:

impl_web is starting. Can you add it to your roster?

The watcher replies once the agent is on its roster, and starts watching it at its next check.

Use the agent’s exact name. If a name on the roster doesn’t match an agent, the watcher may not tell you.

What it looks for

On each check, the watcher reads the terminal of every agent on its roster. Among other things:

• An agent running out of context. By default it warns an agent once it has used about half of its context window, the space the AI tool has for the conversation. It compacts an agent (shrinks its conversation) only when the agent is out of context and has been at rest for two checks in a row.
• An agent that has gone quiet. If an agent said what it would do next and then sits idle for two checks, the watcher reminds it.

It raises the same problem no more than three times.

Permission prompts go to your coordinator

A permission prompt is a question an agent asks before it runs a command or edits a file. When an agent stops at one, Thrum tells your coordinator, which looks at the agent’s terminal and decides. The walkthrough on the homepage shows this: a reviewer stops to ask before it runs git worktree add, and the coordinator checks the reviewer’s terminal and approves it. Answer permission prompts shows how to answer one yourself.

Which tools can run a watcher

A watcher can run in Claude Code or Muse.

If the watcher stops

A watcher doesn’t start again by itself after it stops. Ask your coordinator to restart it.

Free and Pro

A watcher works on Free and Pro.

Next steps

• Answer permission prompts: what happens when an agent stops to ask.
• Grow your team: when to add each kind of agent.