One change, three repos, and the orchestrator steps aside

The second example in the repo, examples/02-cmux-handoff, is the demo we wish every multi-agent framework had to pass. The setup: one API change that has to land across three repositories (todo-api, todo-web, and todo-docs), with one agent working in each.

It starts the way you'd expect: an orchestrator DMs each worker its slice of the change. Classic hub-and-spoke, and COTAL expresses it fine: unicast is one of the three delivery modes, and a supervisor is just a peer that happens to send a lot of DMs.

The interesting part happens midway. The API agent finishes first and realizes the web client needs to know the new response shape. It doesn't report back up to the orchestrator and wait for the message to be relayed. It DMs the web agent directly. The handoff is lateral. The orchestrator finds out the way a good manager does: by watching #general, not by being a bottleneck in every conversation.

Why this matters

In most agent frameworks, topology is baked into the code. If you built hub-and-spoke, every byte flows through the hub forever. The coordinator becomes a context-window tax on the whole system. In COTAL, topology is just a pattern of messages over the same three primitives. Start supervised, go lateral when it's faster, mix both in the same session.

The whole run is observable from a fourth terminal with cotal-ai watch --space demo: every channel post, claim, and handoff in one feed. Try it from the repo; it's four terminals and a coffee.