AWS AI Labs details AECP protocol for multi-agent code generation

In a new research paper, AWS AI Labs introduced the Artifact-Exclusive Communication Protocol (AECP), a system designed to improve coordination among multi-agent coding teams without altering underlying models. In conventional free-form agent setups, shared information functions only as unstructured context that agents can ignore, while harnesses fail to verify whether code implementations align with agreed interfaces. Under AECP, agents communicate exclusively through structured artifacts managed directly by the execution harness.

The harness injects recorded findings whenever an agent opens relevant code, validates implementations against interface commitments, and prompts affected agents to renegotiate agreements when changes occur. Across the Doc2Repo, NL2Repo, and CodeProjectEval benchmarks using models including Opus-4.8 and DeepSeek-V4-Flash, the protocol raised test pass rates by 28.2% and reduced wall time by 16.5%.

The authors also observed security improvements under the protocol. Malicious instructions relayed between agents reached another agent 0% of the time instead of 95%, and execution of those instructions dropped from 40% to 0%. Technical details and experimental results are available in the [AECP paper](https://academy.dair.ai/papers/aecp-artifact-exclusive-communication-protocol-for-multi-agent-code-generation-2610.06481).