Multi Agent Protocol

multi-agent-protocol

**A multi-agent collaboration protocol for running AI agent teams on real work — without losing control of authorization, history, or quality.** Five default participants; four of them hold authority (principal, owner, builder, reviewer) — the orchestrator is an interface, never an authority.

[](LICENSE) [](https://github.com/AIpandadreams/multi-agent-protocol/actions/workflows/mirror-check.yml) [](https://github.com/AIpandadreams/multi-agent-protocol/releases) [](docs/PROTOCOL.md) [](https://www.python.org/downloads/) [](docs/QUICKSTART.md#1-install-the-plugin) [](#trust-properties)

(The release number and the `PROTOCOL vX.Y` stamp in the badges above are **different counters** that move independently — see [PROTOCOL.md § Version spaces](docs/PROTOCOL.md#version-spaces).)

Most multi-agent setups fail the same three ways: an agent quietly does something it was never authorized to do, a session dies and takes its state with it, or agents review their own work and call it done. This protocol is the distillation of a system that ran for months on real production work — every rule in it exists because its absence caused a real defect.

The parties