ARCHER: deterministic multi-agent orchestration lets a self-hosted open-weights model reach 97.8% of frontier accuracy at a quarter the cost

arXiv·medium signalXBlueskyLinkedInCopy link

ARCHER is a test-driven, deterministically orchestrated multi-agent program-synthesis harness that generates auditable verification code from regulatory Codes of Practice for building-compliance checking, addressing the fact that most automated compliance checkers are proprietary and cannot be inspected for whether they captured regulatory intent. Evaluating a taxonomy of six harnesses of increasing agentic sophistication across four backbone models — from frontier third-party APIs down to a fully on-premise open-weights model — ARCHER's deterministic orchestration wins for every backbone, improving mean union accuracy 82% over naive single-pass prompting. The cost-accuracy result is the transferable one: harness design, not model tier, closes most of the gap, making data-sovereign agentic workflows practical.