AMC

A policy is a claim about the next run

A policy tells me what an agent is supposed to do. I need a run to see what it actually did. I built Agent Maturity Compass so an assessment can reflect that difference. A team can inspect the evidence behind a score and decide which control to work on first.

Let the run change the score

AMC assesses maturity from L0 to L5. It weighs observed, attested and self-reported evidence differently. A policy document can show that a team has specified an approval step. Execution evidence can show whether the agent waited for approval before it acted. That distinction matters when the next person has to decide how much authority to give the agent.

I designed the assessment around a loop: diagnose a gap, change a control and run the agent again. A revised score depends on evidence from the new run. AMC covers eight capability families: scoring, adversarial assurance, enforcement, evidence storage, monitoring, compliance mapping, fleet trust and agent identity.

What teams can run today

The MIT-licensed v1.1.1 release is published on npm with CLI, Studio, API and MCP interfaces. I also led the AMC white paper, now at v2.0, published under POLARIS Research Team and AMC Labs. In the v1.2 source, I’m building a native runtime that owns the model loop, tools, approvals, signed sessions and MCP connections. Those runtime features are in development.