← ResourcesGuide · 5 min read

What is GovernanceBench and how do I run it?

GovernanceBench is the open, reproducible benchmark for AI agent governance, published by Agentomy under Apache-2.0. It runs real attack and failure scenarios against any platform, including Agentomy's own, and scores six dimensions. Here is what it measures and how to run it in minutes.

Guide

GovernanceBench measures whether a governance layer does what it claims when an agent misbehaves: whether an unauthorized action is blocked, whether the record of it can be proven, whether the agent can be stopped mid-action, and whether behaviour, override and message governance hold under adversarial scenarios. It is a benchmark of the governance layer itself, not of the agent's task performance, and it is published so that any platform can be held to the same scenarios.

What it measures

  • Six governance dimensions: authorization, auditability, override capability, behavioural integrity, OWASP Agentic coverage and message governance.
  • A core conformance set of 235 real attack and failure scenarios that every implementation is held to, plus optional capability suites for vertical and workflow governance.
  • Reproducibility: the runner versions, the scenario-set hashes and the verified figures behind every published score are in the result manifest at agentomy.com/benchmark-manifest.json, rendered from the verification receipt on every verified change.

How to run it

  • Against any platform: npx governancebench run --target <url>. The runner is an npm package under Apache-2.0; the source is the getagentomy/governancebench repository.
  • Against Agentomy on your own machine: clone the platform, docker compose up, and point the runner at the local endpoint. The evaluator quickstart in the developer docs walks the ten-minute path to a first report.
  • Read the result beside the manifest: a score is a claim about the scenarios it ran, and the manifest names them.
The stronger the number, the stronger the evidence needs to be. The benchmark is open so the evidence can be anyone's.

Where to verify this yourself

  • The benchmark page, with the published results and the FAQ: agentomy.com/governancebench.
  • The published result manifest: agentomy.com/benchmark-manifest.json.
  • The developer docs hub, including the evaluator path: agentomy.com/docs.