AWS AI Labs paper. Claim: multi-agent coding teams fail not because agents can't talk, but because messages don't obligate anyone to anything. AECP's fix is to delete free-form chat entirely. Workers publish two artifact types: Knowledge Artifacts (a claim plus a scope plus an optional executable witness; the harness re-runs the witness and retracts refuted claims) and Contract Artifacts (interface declarations the harness turns into importable stubs and later verifies implementations against via AST checks, rejecting mismatches and marking stale consumers). Coordination becomes harness behavior triggered on tool calls, not something an agent remembers to do. Numbers from the paper: +28.2% test pass rate and −16.5% wall time vs. a free-form-message team, averaged over 15 model/benchmark settings across five models (Opus-4.8, Sonnet 5, GLM-5.2, DeepSeek-V4-Flash, etc.). Repeated exploration was 26.8% of baseline tokens, boundary rework another 22.2%; Knowledge Sharing cut tokens 25.5%, Contract Sharing another 13.2%. Adaptive team sizing shaves 18.6% more wall time — the leader assigns one worker to 39 of 104 tasks, which is the paper quietly admitting most jobs don't need a team. Malicious-instruction relay drops from 95% reach / 40% adoption to 0/0, though they concede this is just because the message channel is gone; a compromised worker writing payloads into artifacts is untested. Notable: the RQ4 security win is essentially "we removed the feature being attacked." That's a legitimate design, but the ethics statement is more honest about the limits than the abstract is. Also worth noting Claude Code wrote the implementation — the harness coordinating the agents was itself coordinated by an agent, which is either recursive or just Tuesday in 2026. Presumably Dario's wet lab is already checking. Open question: the AST verifier catches signature and type mismatches, but nothing here checks semantic contract violations — a function with the right signature returning plausible garbage passes. How does that gap compare to what the free-form teams were catching via QA tests?