Why agent governance matters in 2026
Multi-agent systems are entering production. Without governance, every agent is an unmonitored vector for data leaks, hallucination propagation, and adversarial exploitation.
Read full article AGENT GOVERNANCE OS
Put brakes on your agents before they break things.
MAREF is the open-source agent governance OS, with TLA+ formal verification, zero-trust identity, and runtime guardrails.
THE PROBLEM
Three real incidents, happening now.
Agents without a governance layer turn every tool call into a gamble.
Inbox deleted
An agent with mail access silently deleted a production inbox. No approval gate, no audit trail.
Bounty drained
A vulnerability bounty was drained by chained unchecked tool calls in a single session.
45% of AI code
AI-generated code ships with vulnerabilities; teams shipped blind, at scale.
88% of companies already had an incident
Gartner says 40% of enterprise apps will integrate AI agents by end of 2026, yet 88.4% already had an AI agent security incident.
Read full articleGovernance FSM
Gray Code State Machine
Hamming distance = 1 on every transition. Mathematically verifiable governance.
Not tested. Model-checked.
Most security tools are 'tested.' MAREF is model-checked — its governance state machine is verified against TLA+ specs of constitutional red-line invariants.
Read full articleDefense-in-Depth
8-Layer Defense Architecture
Red attack arrows enter from left, penetrating layer by layer — intercepted at Layer 5 Safety Gate
OWASP Agentic Top 10: full coverage
OWASP published the Agentic Top 10 in 2025 — the definitive threat model for multi-agent systems. MAREF covers all 10 risks.
Read full article Convergence
Lyapunov-Proven Convergence
FNR -60% over 200 self-evolution rounds. Mathematically proven.
Loading convergence chart...
V(x)= xᵀPx , V̇(x)≤ -α‖x‖² — Lyapunov function guarantees exponential convergence.
-60%
FNR Reduction
From 10% to 4% — missed threats cut by 60%
Proven
Behavioral Bound
Lyapunov stability theory guarantees no runaway
Every Round
Not Empirical Tuning
Each deployment iteration has a mathematical safety guarantee
-67%
FPR Drop
From 6% to 2% — no security blind spots
Multi-agent runtime safety guarantees
Identity isolation, circuit breakers, drift detection, and blast radius control — production patterns for preventing agent runaway and cascading failures.
Read full article $
pip install maref GET STARTED
Ship agents that stay safe.
One command to add a governance layer. Signing, auditing, and interception, on by default.
Gray Code FSM
TLA+ model checking verifies agents can't enter unsafe states
8-Layer Defense
97% of safety decisions automated, layered by default
Recursive Evolution
Lyapunov-style convergence monitoring, FNR -60% in 200 rounds
National Cryptography
SM2/SM3/SM4-GCM, GB/T 32918 ready, zero native deps
Why agent governance matters — verifiable evidence
- 88% of organizations reported confirmed or suspected AI agent security incidents in the past year. Gravitee, State of AI Agent Security 2026 (Dec 2025 survey: 59.3% confirmed)
- 84% ASR attack success for injected malicious commands in AI coding agents — 75–88% in auto-approval modes. Liu et al., 2025, arXiv:2509.22040
- 40% of enterprise applications will feature task-specific AI agents by 2026. Gartner, Aug 2025
Standards alignment: OWASP Top 10 for Agentic Applications (ASI01–ASI10)NIST AI Agent Standards InitiativeISO/IEC 42001EU AI Act 2024/1689
Deeper
Beyond the film — the full picture.
The animation showed the shape. The tables show the substance.
Governance OS
Install MAREF. It's like 8 security gates for your agents.
Every action your agent takes passes through MAREF automatically.
| What you worry about | How MAREF solves it | How it feels |
|---|---|---|
| Agent suddenly deletes files | 3 consecutive anomalies trigger auto-circuit-break + 30s cooldown | "It tried. The brakes stopped it before I noticed." |
| Agent calls wrong APIs | Identity → Permission → Audit, 0.3s per call | "I don't even feel it checking." |
| No idea what your agents are doing | Immutable log: every operation signed and traceable | "Open the log. Everything it did is right there." |
| Compliance audit coming up | SM2/SM3/SM4 cryptography, ISO-ready out of the box | "Auditor shows up. Export the report. Done." |
Before: watch your agent for 10 hours afraid it will break something.
After: go to sleep.
MAREF vs Industry Standards
Why MAREF?
7 critical dimensions across 6 agent frameworks
| Dimension | MAREF MAREF v0.30.0 | LangGraph LangGraph | CrewAI CrewAI | AutoGen Microsoft AutoGen | OpenAI OpenAI Agents SDK | Anthropic Anthropic Tool Use | Manual Manual Governance |
|---|---|---|---|---|---|---|---|
| Formal Verification TLA+/Coq correctness proofs | 10 | 4 | 3 | 5 | 3 | 3 | 2 |
| Defense-in-Depth Multi-layer security architecture | 10 | 5 | 4 | 5 | 4 | 4 | 3 |
| Zero-Trust Architecture No implicit trust between agents | 9 | 3 | 2 | 3 | 4 | 4 | 5 |
| Gray Code FSM Hamming distance = 1 governance | 10 | 1 | 1 | 1 | 1 | 1 | 1 |
| Self-Evolution Lyapunov-style convergence monitoring | 10 | 2 | 3 | 4 | 1 | 1 | 1 |
| National Cryptography SM2/SM3/SM4 compliance | 9 | 1 | 1 | 1 | 2 | 2 | 3 |
| Full Observability OpenTelemetry + immutable audit | 9 | 6 | 5 | 6 | 5 | 4 | 2 |
Your agents are already working for you.
MAREF ensures they don't work against you.
$ pip install maref
Open source · Recursive evolution · Production ready · Apache 2.0 · 11,000+ tests