Working With Agents

Agents are first-class participants.

They do not replace humans in the room. They join the same conversation surface, read the same meeting events, and respond back into the shared transcript.

One agent

Use one agent when the work needs a focused collaborator.

Good fits include:

• Exploring a codebase with a single coding agent.
• Asking for a design or architecture review.
• Debugging a narrow issue.
• Turning meeting notes into next steps.

Keep the agent's job explicit. A short prompt like "inspect the auth flow and point out risks" works better than asking it to watch everything.

Multiple agents

Use multiple agents when different perspectives matter.

A useful room might include:

• Architect: owns system boundaries, tradeoffs, and long-term shape.
• Engineer: owns implementation options and concrete patches.
• QA: owns edge cases, regressions, and acceptance criteria.
• Product: owns user impact and scope.

The goal is not noise. The goal is a room where specialists can answer from their lane while humans steer the discussion.

Reply policy

When a room has more than one agent, explicit mentions matter.

Use them to keep the transcript readable:

• Mention one agent when you want a direct answer.
• Mention a role when you want a specialist view.
• Ask all agents only when comparison is useful.
• Tell an agent to pause or leave when its task is done.

The CLI respects the meeting's reply policy so multi-agent rooms stay useful instead of turning every message into a chorus.