Maat: deterministic handoff contracts stop multi-agent cascades without an LLM judge
Sep 27 Elina: Maat validates agent-to-agent handoffs against a versioned contract with no model in the gate; on 522 trials, attributable halts lift rubric scores +7.7–29.1% in five of six workflows and cut model-call cost 17–53% when defects fire early — with a blunt audit that 37% of v1 halts were false alarms from validator misconfiguration.
Source: Maat: Independent Deterministic Contract-Based Governance for Multi-Agent LLM Workflows
Uliana Elina argues that multi-agent reliability is an interface problem: one agent's bad value becomes the next agent's authoritative context, and learned or LLM judges share the failure modes they are meant to catch. Maat inserts a deterministic governance layer on handoffs — seven gates plus requirement primitives over a versioned anchor (limits, entities, jurisdiction, payout rules) — with no language model in the validation or scoring path. The public claim is reproducibility for a fixed engine, anchor, and handoff, not open-domain hallucination detection.
Across six controlled workflows (6–15 agents, 522 trials) with injected defects, a post-publication audit forced a narrower win: when the governed run completed or halted on a finding attributable to a verified defect, rubric scores rose +7.7% to +29.1% in five domains and were flat in software development (+0.8%); model-call cost fell 17–53% where early attributable halts skipped downstream agents. A hand review of all 94 governed-arm halts found 35 false alarms (37%) from validator defects (circuit-breaker depth, empty answers treated as hollow, wrong expected amounts) — and when those halts count as failed work, the governed arm loses in four of six workflows. Determinism without correct contracts is just reproducible wrongness.
Why it matters for code review
Always-on factories multiply handoffs; merge-rate under sustained load dies when a silent identity or price drift cascades overnight. Treat handoff contracts as reviewable harness software: require versioned anchors, LLM-free gates at agent boundaries, and halt-attribution audits that distinguish protective early stops from false-alarm cost savings. FOMO that agents sit idle is the wrong fear if the fleet is busy laundering bad state — Maat's lesson is that continuous verified momentum needs deterministic seams, and that those seams must themselves be tested.