Skip to content

Cognition Protocols

Cognition Protocols occupy the agent cognitive layer, L9, above the L8 agent transports.

L8 standardises message envelopes, performatives and interaction patterns, but does not define a verifiable shared meaning for the content those messages carry, nor primitives for coordination and consensus across a population of agents.

Cognition Protocols are open, measurable specifications that fill that gap — each one satisfies three requirements by construction: the exchange converges provably, emits a certified record of what was achieved, and runs within a cost bound declared at the start.

Common Principles

Each Cognition Protocol satisfies the same three guarantees by construction — convergence, certified benefit, and bounded cost — rather than leaving them as emergent properties to be observed. See Common Principles for the full requirements.

Getting Started

Adding a protocol to an existing multi-agent system requires no application code changes. See the getting started guide for how skills are packaged and deployed across different agent harnesses.

Protocols

The current protocols are:

  • ACCORD — addresses absence of shared intent: misaligned terms, incompatible task models, and outputs produced without an agreed verification step. Establishes who is participating, what the problem is, what the terms mean, and how correctness is checked — before work begins.
  • CONCORD — addresses unreconciled disagreement: divergent-but-correct specialist outputs that converge on an action without consensus. Reconciles differing positions into a certified decision.

For full design details see the Protocols section.

Evaluation

A protocol specification is only as useful as its evidence. Each Cognition Protocol is evaluated against independent dimensions — convergence, correctness, cost, and benefit — so claims about behaviour are measurable, not asserted. See the evaluation overview for results and methodology.

Guided tour

See ACCORD and CONCORD work without running anything: a step-by-step tour of recorded AgentBridge runs, with and without a protocol. See the guided tour.

Demos

Recorded end-to-end demos across SRE triage, e-commerce, and multi-agent coding teams — with and without cognition protocols. See the demos section.

Get Involved

Specifications, skill packages, evaluation harness, and demo recordings are all open. Issues and contributions welcome on GitHub.

Background reading