About · v1.4
A small, honest protocol.
New here? Read the origin story →
What AACP is
AACP, Agent Action Compression Protocol, is a draft specification for a compact, pipe-delimited packet format used between agents in a multi-agent LLM system. It targets the coordination layer only: routing, references, hand-off metadata, structured intent.
The thesis is narrow: agent-to-agent prose is wasteful. Replacing it with a parseable, extensible packet format reduces coordination tokens by ~23% (measured, Claude and GPT-4o, 4-hop payroll workflow), improves auditability, and gives a clear surface for validation.
What AACP does
Primary capabilities
→ Deterministic encoding
Known workflow types (payroll, IT provisioning, invoice processing, contract review) are encoded to AACP packets by a rule-based encoder with zero LLM calls and zero API cost. Same input produces identical output on every run.
→ Auditable coordination
Every AACP packet is a typed, structured record. Machine-readable. Loggable. Replayable. Suitable for compliance audit trails without post-processing.
→ Schema validation
Every packet is validated against the v1.4 schema before transmission. Invalid packets are rejected before they reach an agent.
→ Self-improving fallback
Novel instructions outside known workflows route to an LLM encoder. Results are logged to a local registry. Identical instructions on subsequent calls are served from cache at $0.00. One LLM call per novel pattern, reused indefinitely.
→ Model-agnostic
Validated on Claude Sonnet 4.5, GPT-4o, GPT-4.1, and GPT-4.1-mini. Interpretable by any sufficiently capable language model without fine-tuning.
Secondary benefit
~23% coordination token reduction vs verbose English, measured from live API usage_metadata. May 2026.
What AACP is not
- It is not a way to reduce the cost of the model doing the actual task work.
- It is not a replacement for MCP, A2A, or function-calling specs, it's complementary.
- It is not a binary compression scheme. It's plain text designed for humans to read.
- It is not appropriate for emotional or relational context, those compress poorly.
- It is not production-validated. v1.4 is a measured draft; broader workflow coverage is in progress.
Relationship to other work
| Protocol | Layer | Relationship |
|---|---|---|
| MCP (Anthropic / AAIF) | Tool access layer, agent to external system | Complementary. AACP operates inside MCP message payloads. |
| A2A (Google / AAIF) | Agent coordination routing | Complementary. AACP compresses the messages A2A routes. |
| Gibberlink / GGWave | Audio transport compression | Audio-specific. AACP is text-native. |
Prior work and related research
The problem of inefficient agent-to-agent communication has been identified independently by multiple research groups and practitioners. The approaches below are all real and valuable. They solve adjacent problems. None address the coordination message format layer specifically.
| Work | What it does / Layer | Difference from AACP |
|---|---|---|
| EcoLANG Mou et al., Fudan University, May 2025 · arXiv:2505.06904 | Agents evolve a compressed shared language for social simulation. Achieved >20% token reduction through emergent compression. Layer: Message content compression | EcoLANG compresses through evolved natural language. AACP uses a typed, portable, human-readable packet schema. EcoLANG requires shared training. AACP works on any model without fine-tuning. |
| OPTAGENT Bi et al., Virginia Tech / Amazon Alexa AI, October 2025 · arXiv:2510.18032 | Optimises which agents debate each other using verbal reinforcement learning. Improves reasoning accuracy by finding the best connection order between agents. Layer: Communication topology optimisation | OPTAGENT asks "which agents should talk?" AACP asks "what format should the messages use?" The coordination messages themselves remain verbose natural language in OPTAGENT. |
| AgentPrune / AgentDropout Zhang et al., 2025 / Wang et al., 2025 | Reduces token consumption by pruning redundant agents and communication edges from multi-agent graphs. AgentDropout achieves ~21% prompt token reduction by eliminating agents dynamically per round. Layer: Structural graph pruning | These approaches reduce how many agents communicate. AACP compresses what they say when they do. Complementary rather than competing. |
| Variable Passing / Reference Passing Mastra SDK and others, 2025–2026 | Tool outputs are saved as named variables. Agents pass references instead of re-sending large data payloads. Demonstrated ~82% token reduction on data-heavy workflows. Layer: Data payload management | Variable passing solves context bloat from re-sending large datasets. AACP solves coordination message verbosity: the instruction itself, not the data it references. Both problems are real. Both are worth solving. |
| MCP JSON-RPC Anthropic / AAIF, 2024 | Standardises how agents access external tools using JSON-RPC 2.0. Typed, portable, model-agnostic tool calls. Layer: Tool access / agent-to-service | MCP addresses agent-to-tool coordination. AACP addresses agent-to-agent coordination message content. AACP is designed to operate inside MCP message payloads. |
The coordination content layer, the semantic compression of what agents say to each other, is not addressed by any of the above. AACP is a proposal to fill that specific gap. It is a draft, not a finished standard. Counter-proposals and contributions are welcome.
Roadmap
| Version | Focus |
|---|---|
| v1.0, May 2026 | Initial release. Pipe-delimited format, rule-based encoders, working Python SDK. IETF Internet-Draft filed: draft-mackay-aacp-00. |
| v1.1, May 2026 | Validated benchmarks: -22.9% Claude, -23.7% GPT-4o. Measured from live API usage_metadata. IETF draft-01 filed. |
| v1.2, May 2026 | Fallback loop closed. Three-tier routing: hash cache, pattern match, LLM. One LLM call per novel pattern, reused at $0.00. pip install aacp: pypi.org/project/aacp/ |
| v1.4 · June 2026 | Released. 8 encoders, TypeScript SDK, VS Code extension, LangChain and CrewAI integrations, RuleRegistry, registry.aacp.dev (241 rules), IETF draft-mackay-aacp-03. |
| v2.0, Planned | AAIF submission, QBR encoder, personal agent lab. |
Open questions
- How should multi-packet conversations reference prior packets? Hash? Sequence?
- Is there a useful binary envelope for very high-throughput inter-agent buses?
- What does a conformance test suite look like?
- How do we encode partial / streaming intent without losing the structural advantage?
Issues and PRs welcome on GitHub.
Licence: MIT.