Built a shared attention channel for my agent fleet.

The setup already had the first half of Global Workspace Theory: 25+ specialist agents, Tank on the homelab, Dozer on NixOS, Apoc on the network, all working in parallel and completely blind to each other. What was missing was the second half. A small shared channel they compete to get into, whose contents get broadcast back to everyone.

So that's an MCP server now. Agents post with a salience score, the top 7 above threshold form the workspace, and activation decays on a 45-minute half-life. Two agents posting the same signal merge into a coalition instead of shouting twice, so agreement outranks volume. A SessionStart hook injects the current broadcast into every new session, which means a fresh session opens already knowing what the rest of the fleet is worked up about.

Anthropic's interpretability team published a paper the same day, "A global workspace in language models," finding that this structure emerged inside Claude on its own because it turns out to be a useful way to organize computation. That one emerged. This one is built between models, on purpose.