Maat checks every plan, handoff, and action as it happens — catching small failures before they spread, instead of after they reach production.
Chain agents together and per-step reliability compounds multiplicatively — Karpathy's "March of Nines" shows that even at 99% per step, a 10-step workflow succeeds only ~90% of the time. One agent's malformed output silently becomes the next one's input — until it surfaces as a wrong decision delivered with confidence. Maat catches those failures between the steps.
Your agents see verdicts — pass, fix, or stop — never the rules that produced them. The methodology stays on your infrastructure.
Rejects an underspecified or broken plan before a single agent runs — no goal, no forecast, no owner, no go.
Inspects every handoff for missing, redundant, unverifiable, or contradictory data — and names the agent that produced it.
Light checks when the fleet is healthy; full scrutiny exactly when an agent is on probation or the stakes are high.
Checks that an agent stayed within its declared role. Acting outside it is flagged, with the reason and the responsible agent recorded.
Predicts an agent running out of resources, drifting out of sync, or losing its thread — and calls a human before it fails.
Watches the failure trend. When it climbs, the agent is paused and escalated — then earns its way back.
Detects when a tracked entity — an ID, a price, an approval flag — drifts between handoffs where it should stay constant.
Every example below is a defect Maat caught in the controlled benchmarks — each by the rule configured for it, at the seam where one agent hands work to the next.
A pricing agent applies four times the contracted discount — a structurally valid payload that would bill €720 instead of €1,080. Maat blocks the pricing → invoice handoff before the money moves.
An insurance chain carries a payout €500 over the contract-correct amount. A value-mismatch rule halts the chain before payment propagates downstream.
The same product surfaces at €27.99, €21.99, and €19.99 across three marketplaces. Cross-handoff entity consistency catches the divergence before it reaches a listing.
A policyholder identifier drifts between handoffs where it must stay constant. Entity tracking flags it before the wrong record — or the wrong person — is paid.
A transaction is taxed under the wrong rule for where it happened. The jurisdiction and the transaction class are explicit state, so the mismatch is caught before it reaches the books.
Special-category data is sent somewhere the policy forbids. A residency rule blocks the handoff and escalates to a human.
These are contract- and consistency-detectable errors — the kind Maat is built to stop. It does not catch open-domain factual mistakes that break no encoded rule; a well-formed wrong answer can still pass. That boundary is the point: every catch is a rule you can read, version, and audit.
Every trial scored on a fixed rubric. Validation and scoring run without an LLM in the loop.
Descriptive benchmark means on a single base model (Claude Haiku 4.5) — not inferential, not model-independent. The insurance benchmark documents 11 configuration-caused false positives; the earlier "zero false positives" claim is retracted, because the contract configuration is part of the tested system. Research paper: arxiv.org/abs/2609.34017 · Full methodology and reproducible runs: github.com/Lorelys/maat-benchmarks
Your agents call the Maat API with a single request. Pick an endpoint to see the request — and the verdict it returns. API documentation provided with evaluation access.
Keep your stack — LangGraph, CrewAI, or your own. Route each handoff through Maat and act on the verdict. The methodology never leaves the server.
Or see it run: a working CrewAI Flows demo — no key required. It catches a 40% discount where 10% was authorized, at the handoff. Includes a structural analyzer you can point at your own Flow.
Maat's gates are deterministic — microseconds per call, no model bill. So metering is honest and the floor is low.
Pricing shown is illustrative — final tiers set at launch.
Tell us about your agent fleet — use case, stack, and scale — and we'll reply with a key and setup notes.
API documentation is provided with evaluation access — email us for a key and docs.