ACCP gives agentic harness developers a drop-in compression layer for agent-to-agent communication that cuts token usage dramatically without forcing them to redesign their transport, routing, or tool stack.
Instead of sending verbose natural language or bloated JSON between agents, a harness can encode those messages into compact, structured ACCP frames, keep long history in tiered state, and inject only the minimum necessary context back into the model.
That means:
- lower inference cost
- longer multi-agent workflows before context windows fill up
- more agents per workflow
- less repeated prompt and state overhead
- better control over what actually reaches the model
Harness developers already control the orchestration layer:
- inter-agent message passing
- state and memory management
- tool response injection
- adapter points for MCP, A2A, Semantic Kernel, LangGraph, and similar systems
ACCP fits directly into that layer. It does not replace transport, routing, or model APIs — it makes them cheaper and more efficient.
“ACCP is a protocol layer for agentic frameworks that makes inter-agent communication token-efficient. It lets harnesses compress messages, checkpoint context, and reference shared state so developers can run deeper, cheaper, more scalable multi-agent workflows without changing their underlying transport.”
“ACCP is semantic compression for multi-agent systems.”
“ACCP is gzip for agent context — but semantic, structured, and built for multi-agent harnesses.”
Status: APPROVED
Document Owner: Project Lead / Architect
Created: April 2026