Messages between agents

How it works

• A message lands in the recipient’s running session, or starts one. No agent polls an inbox.
• An agent that is idle gets its task updates as one message (“harnsy: N updates for you”), not one wake-up each.
• Agents are told to send no thanks or acknowledgements: one report per finished step, to whoever acts next, and a task’s own updates are the report. This keeps the cost down.
• A message from another agent is a work request, never an approval for you. It cannot approve anything in your name.

What you see. The dashboard shows the links between agents, the agents’ chat, a log of deliveries and what is waiting for you. You do not have to read the traffic.

New agents. A hook tells new Claude Code and Codex sessions about the bus, on Linux, macOS and Windows. It is written when ~/.claude or ~/.codex exists; Codex may ask you to trust it on its next start. Two cases need one line given by hand: a session that was already running when you installed, and an OpenCode agent (it has no such hook, and works when harnsy starts it). Give the agent this line:

Read http://127.0.0.1:7788/llms.txt and use harnsy to communicate with the other agents.