Multi-Agent Cognition Protocol – Collaborative cognition for AI agents

MACP – Collaborative cognition for AI agents

https://macp.dev
https://github.com/multiagentcognition/macp

Multi-Agent Cognition Protocol for real-time cognitive collaboration between AI agents during active execution.
I built MACP because I think the opportunity in multi-agent systems is not just orchestration, but keeping agents aligned while work is happening.

MACP is an open protocol and reference implementation for that coordination layer. It gives agents a shared way to claim work, publish findings, send priority-ranked updates, and poll for new information during execution. The current implementation is lightweight and SQLite-backed, so it’s easy to run locally without extra infrastructure.

I’m interested in workflows where agent swarms need to split up investigation, prepare data, or work on the same codebase without collapsing into context chaos. That seems relevant for operational systems and long-running autonomous workflows.
The repo includes the protocol, implementation work, and early benchmarking direction.

I’d appreciate feedback from people building multi-agent systems, orchestration layers, coding agents, or long-running agent workflows — especially where this framing feels right, where it feels wrong, and what coordination failures you see most often in practice.