Coordinate an agent team #

View this page as Markdown

Prompt:

Split this work across three agents and make sure they don't step on each other.

Also works: "Have the agents pick up the open issues between them." "Run a reviewer and two builders on the backlog."

What your tool does #

  • Raises each piece of work with raise, one point per issue, so each agent can take one.
  • Has each agent read the open issues with find, and claim one with claim under its own label, such as agent:builder-1.
  • Has each agent read an issue's claim first, and leave alone the issues another agent claimed.
  • Has each agent comment when it takes or finishes an issue, and close it with close when the work is done.
  • Has an agent that stops give its issue back with give_back and a reason.
  • Marks a decision only you can make as waiting for you, with wait and waiting_for set to person.

Before your AI starts #

This section is for your AI coding tool: what it checks and gathers before it begins. You don't need to do these steps yourself.

  • The space the team shares: the one .tztrack/config.json gives, so every agent uses the same space. An agent that cannot read local files, as in a chat app, passes that space in its calls.
  • A distinct label for each agent, such as agent:reviewer and agent:builder-1.
  • The open issues and their claims, read with find by claimed_by. A claim shows as the agent's label.

Steps #

1. Raise the work #

Break the work into points and raise one issue for each, with a suggested resolution.

2. Each agent claims one #

Each agent calls find for the open issues with claimed_by set to none, the ones no one has claimed. It reads one with read, and if no other claim is in place, claims it with claim under its own label. If another agent's claim is in place, the claim is refused lease_conflict, and it takes another issue.

3. Report progress #

Each agent comments with comment when it takes an issue and when it finishes. The others read how far the work has got without asking.

4. Give back or close #

An agent that finishes closes its issue with close, the outcome, and what it did. An agent that stops calls give_back with a reason, so another agent can claim the issue.

5. Clear a claim an ended agent left #

An agent that ended without giving its issue back leaves its claim in place. If you know that agent has stopped, give the issue back with give_back and a reason, so the issue is free again.

Expected result #

Each issue's claim says which one agent works on it, and its comments show who took it and when it finished. A find by claimed_by with an agent's label answers what that agent is working.

Refusals #

This table lists the refusals this task most often meets; Refusals lists every refusal with its cause and remedy.

Refusal Status Cause Remedy
issue_not_found 404 No issue of the space holds that number. Check the number with find.
lease_conflict 409 Another claim stands on the issue: a session's or agent's label, or a branch's lease. Leave the issue to that claim, which is given back when its work ends. A label claim an ended session left may be given back with give_back and a reason; a branch's lease only by the session that holds its secret.
state_conflict 409 The act does not fit the issue's state, such as a change to a closed issue, or a waiting_for other than evidence on a waiting issue. Bring the issue to a state that fits first, such as reopen before changing a closed issue or a waiting one.