Maat Layer is a resilience API for multi-agent AI systems. It judges every plan, handoff, and risky action before it proceeds and returns pass, fix, or stop. The methodology stays private on your infrastructure; agents only see verdicts.

What are the seven gates in Maat Layer?

Plan validation; handoff and information-defect validation; validation depth; integrity and role adherence; agent health and survival; circuit breaker; and cross-handoff entity consistency — detecting when a tracked entity (an ID, a price, an approval flag) drifts between handoffs where it should stay constant. Together they catch structural, drift, runaway, and reliability failures in agent fleets.

How is Maat Layer different from guardrails?

Guardrails often filter single-model outputs. Maat Layer governs fleet structure: plan shape, inter-agent handoffs, action risk, health trends, and audit trails — with findings attributed to the producing agent.

Does Maat Layer work with LangGraph or CrewAI?

Yes. Maat is orchestrator-agnostic. Wrap handoffs and actions with HTTP calls to the Maat API and enforce verdicts in your existing LangGraph, CrewAI, or custom workflow.

What does Maat Layer not do?

It does not make a single agent smarter, stop all hallucinations, or provide security against prompt injection. It judges whether work is structurally sound and stops defects from spreading — your orchestrator still executes.

Where are the Maat Layer API docs?

API documentation is provided with evaluation access. Request access at hello@synwe.ai.

Maat Layer — resilience engineering for AI agent fleets

Why fleets fail

Small errors don't stay small. They compound.

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.

~90%

success for a 10-step workflow at 99% per-step reliability — Karpathy, "March of Nines"

37%

of multi-agent failures are inter-agent misalignment at the handoff — MAST (Cemri et al., 2025, UC Berkeley)

Seven gates, one sealed engine

Glass box on inputs and outputs. Sealed engine in between.

Your agents see verdicts — pass, fix, or stop — never the rules that produced them. The methodology stays on your infrastructure.

GATE 01

Plan validation

Rejects an underspecified or broken plan before a single agent runs — no goal, no forecast, no owner, no go.

GATE 02

Handoff / information-defect validation

Inspects every handoff for missing, redundant, unverifiable, or contradictory data — and names the agent that produced it.

GATE 03

Validation depth

Light checks when the fleet is healthy; full scrutiny exactly when an agent is on probation or the stakes are high.

GATE 04

Integrity & role adherence

Checks that an agent stayed within its declared role. Acting outside it is flagged, with the reason and the responsible agent recorded.

GATE 05

Agent health / survival

Predicts an agent running out of resources, drifting out of sync, or losing its thread — and calls a human before it fails.

GATE 06

Circuit breaker

Watches the failure trend. When it climbs, the agent is paused and escalated — then earns its way back.

GATE 07

Cross-handoff entity consistency

Detects when a tracked entity — an ID, a price, an approval flag — drifts between handoffs where it should stay constant.

Results

Measured results from six controlled benchmarks — deterministic scoring, no LLM in the validation or scoring path.

Every trial scored on a fixed rubric. Validation and scoring run without an LLM in the loop.

+3.8–26.5%

correctness improved on every benchmark, on a deterministic rubric

−50%

wasted compute on workflows where defective chains were halted early

100%

VAT/tax misclassification caught on every trial where it occurred

clean catches on the insurance benchmark, each by the intended primitive — with every false positive documented, not hidden

Full methodology, scenarios, and reproducible runs: github.com/Lorelys/maat-benchmarks

The API

One call. One verdict.

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.

Request

Maat responds

✕ REFUSE accepted: false depth: "thorough" findings • INFO_UNRELIABLE: "no provenance; cannot verify" • owner: "agent.market_analyzer" note: high-risk handoff held to strict checks

Connect in minutes

One wrapper per agent. Nothing else changes.

Keep your stack — LangGraph, CrewAI, or your own. Route each handoff through Maat and act on the verdict. The methodology never leaves the server.

# before your agent's output reaches the next agent, weigh it
verdict = requests.post(
    "https://api.synwe.ai/v1/handoffs/validate",
    headers={"Authorization": f"Bearer {MAAT_KEY}"},
    json={"plan_id": plan, "handoff_id": step, "produced": output},
).json()

if not verdict["accepted"]:
    # a defective message never reaches the downstream agent
    repair(verdict["findings"])     # Maat tells you what, where, and who

Pricing

Pay for judgment, not idle time.

Maat's gates are deterministic — microseconds per call, no model bill. So metering is honest and the floor is low.

Feather

for builders & trials

All seven gates
Evaluation access
Self-hosted or our sandbox
Community support

Request evaluation access

Scale

usage / verdict

for production fleets

Unlimited verdicts, metered
Per-customer API keys
Persistent audit log
Works with your existing orchestrator (LangGraph, CrewAI, custom)
Priority support

Request access

Sovereign

custom

your infrastructure

Deploy in your own cloud
Methodology stays wholly yours
Custom role charters
SSO, compliance, SLA

Talk to us

Pricing shown is illustrative — final tiers set at launch.

Get started

Request API access

Tell us about your agent fleet — use case, stack, and scale — and we'll reply with a key and setup notes.

Email hello@synwe.ai → Request docs access →

API documentation is provided with evaluation access — email us for a key and docs.