Multi-Agent Coordination Protocol

The era of capable AI agents is here. The missing piece is getting them to converge on a single outcome — reliably, every time, under pressure. MACP is the open protocol that makes multi-agent coordination a first-class primitive.

v1.0.0-draft · RFC-driven · Open Standard
The Challenge

Intelligence is abundant. Coherence is not.

Today's multi-agent systems are built on communication primitives — agents call agents, planners aggregate responses, retry logic smooths uncertainty. From the outside it looks coordinated. But beneath the surface, convergence is implicit. There is rarely a declared boundary where coordination begins, a lifecycle that guarantees it ends, or a structural guarantee that two identical runs unfold identically.

“Agents can talk to each other freely — but those messages never commit anyone. Real decisions only count when they happen inside a Coordination Session.”

— the core rule from RFC-MACP-0001

This single invariant drives the entire protocol design. MACP doesn't define decision theory, governance policy, or domain logic — those belong to Coordination Modes layered above core. What it does define is the structure that makes coordination auditable, replayable, and provably convergent.

Sessions

Every coordination has a beginning, a lifecycle, and a binding outcome. No more emergent convergence — MACP gives multi-agent interactions the same transactional integrity that databases gave to data.

Modes

Five standards-track coordination patterns — Decision, Proposal, Task, Handoff, Quorum — so agents don't just talk, they converge. Each mode defines exactly how participants reach a binding outcome.

Replay

Every session can be replayed bit-for-bit. When a coordination needs auditing months later, you can reconstruct exactly what each agent saw, said, and decided — deterministic by design.

Distributed computing solved this problem for data with transactions and consensus. MACP solves it for AI with coordination modes, session lifecycles, and replay integrity.

What MACP Is Made Of

Agents (backed by LLMs) enter a Session, coordinate under a Mode, and emit a binding Decision.

Participants
Agent A
LLM
Agent B
LLM
Agent C
LLM
Coordination Session

A bounded context where participants exchange typed messages under a chosen mode.

Mode
DecisionProposalTaskHandoffQuorum
Decision Result
Binding · Auditable · Replayable

Explore the Ecosystem

Everything you need to understand, build with, and operate multi-agent coordination — from specification to production.