Agent governance,
rethought from first principles.
From 30 years behind a steadicam on Hollywood sets to building safety systems for AI agents — because when things move fast, safety isn't optional.
MAREF makes multi-agent AI systems measurably safe. Not "we hope it's safe." Governed — with TLA+ model checking against constitutional red-line invariants, Lyapunov-style convergence monitoring, and cryptographic audit trails.
The story so far
- v0.30 2026 · 05
First GA baseline
The governance runtime goes from concept to a runnable, testable system — with a real, signed audit trail.
- v0.34 2026 · 06
Proved, not tested
TLA+ model checking against constitutional red-line invariants, plus G1–G5 cross-verification on every safety claim.
- v0.43 2026 · 07
Open-sourced
Apache-2.0 release. Recursive self-evolution reaches GA on real metric-driven loops, with Lyapunov-style convergence monitoring.
- v0.54 2026 · 08
Production governance loop
Enterprise value loop, governance gap closure, and cost guardrails — the governance OS closes the loop.
Mission
MAREF makes multi-agent AI systems measurably safe. Not "we hope it's safe." Not "our red team didn't find anything." Governed — with TLA+ model checking against constitutional red-line invariants, Lyapunov-style convergence monitoring, and cryptographic audit trails.
We believe that as agents become infrastructure — like databases, networks, and DNS — they need infrastructure-grade governance. Not a bolted-on security scanner. A constitutional layer that every agent, every tool call, every decision passes through.
Why governance-first?
Because the data is clear: AvePoint's State of AI 2026 report found that 88.4% of organizations already had an AI agent security incident. 94% don't have a mature strategy. Companies spend 17x more on "AI-powered security" than on "securing AI itself."
The market is building agents faster than safety can keep up. MAREF closes that gap — not by slowing development, but by putting guardrails in place so teams can move fast without breaking things.
What makes MAREF different
TLA+ model checking
Not "we tested it." The governance state machine is model-checked against TLA+ specs of its constitutional red-line invariants.
8-layer defense architecture
From prompt injection to rogue agent detection. Defense in depth isn't a feature; it's the architecture.
Recursive self-evolution
Red-blue adversarial evolution. Defenses harden over time, with Lyapunov-style convergence monitoring of error metrics.
Zero-trust identity per agent
Ed25519 signing for every decision. Every action is cryptographically attributable.
Project status
MAREF is in active development, released under Apache 2.0. Available on GitHub at github.com/maref-org/maref.
Contact
Technical support: [email protected]