# AACP, Agent Action Compression Protocol

> A measured, draft protocol for structured agent-to-agent coordination in multi-agent LLM systems. Open spec, MIT licensed, v1.1-draft.

AACP replaces wasteful agent-to-agent prose with a compact, pipe-delimited
packet format (TASK|DOM|return:AGENT|p:1|aacp:1.1|...) for the coordination
layer only. Measured ~23% coordination-token reduction on a 4-hop payroll
workflow against live APIs (-22.9% Claude Sonnet, -23.7% GPT-4o, May 2026).
Task tokens are unchanged.

## Pages

- [Home](/): Overview, measured results, interactive calculator.
- [Spec](/spec): v1.1-draft pipe-delimited packet format, required fields, semantics.
- [Encoder](/encoder): Convert structured intent into an AACP packet, including rule-based encoders.
- [Decoder](/decoder): Expand an AACP packet into structured English.
- [Validator](/validator): Check a packet against the v1.1-draft spec.
- [Roles](/roles): Workflow comparisons, verbose prose vs AACP coordination.
- [About](/about): Scope, what AACP is not, related work, roadmap, open questions.

## Optional

- [GitHub repo](https://github.com/MackayAndrew/aacp): Source, issues, PRs.