Skip to content

agentgovernance.dev

AI agents fail in ways monitoring can't catch. This tool detects them.

Built from 11 real incidents operating a governed agent fleet. 8 behavioral patterns documented. Phase-gated execution with human approval at every transition.

Works with CrewAI, LangGraph, AutoGen, or your own agent stack.

npm install @agentgovernance/sdk

import { AgentGovernance } from '@agentgovernance/sdk'

const client = new AgentGovernance({ apiKey: 'ag_...' })

// Register an agent
const agent = await client.registerAgent({
  name: 'my-agent',
  framework: 'crewai',
})

// Assess output for behavioral patterns
const result = await client.assess({
  text: agentOutput,
  agentId: agent.id,
})
// → { patternsDetected: 2, highestSeverity: 'P2' }

What This Does

The governance API provides three capabilities:

  1. Phase-gated lifecycle — PLAN → BUILD → REVIEW → REMEDIATE → APPROVE. No work advances without explicit human approval at each gate.

  2. Behavioral pattern detection — 8 documented failure modes (BP-001 through BP-008) that are invisible to observability tools. The API scans agent output for these patterns.

  3. Audit trail — Every action, gate transition, and pattern detection is recorded with actor identity and evidence.

What This Doesn't Do

  • It doesn't replace observability platforms. It governs what they monitor.
  • It doesn't enforce security. It governs the operational lifecycle of agent work.
  • It doesn't assess model quality. It detects behavioral failure patterns in agent output.

Published Methodology

The behavioral pattern taxonomy and governance methodology are published at aiagentgovernance.org under CC BY 4.0. This tool is the reference implementation.

License

Apache 2.0

Maintained by John McCormick. Questions or feedback — use the contact form.