agentgovernance

AI agents fail in specific, documentable ways. Most teams discover this after the fact.

This is a governance API for AI agents. It detects behavioral failure patterns in agent output, enforces a phase-gated execution lifecycle with human approval at every transition, and records everything to an audit trail.

Where this came from

Every pattern in this library was documented from a real failure — not threat modeling. The taxonomy was built from 14 incidents operating a governed multi-agent system in production, observing how standard incident frameworks could not classify what was failing or why.

The full governance methodology is published at aiagentgovernance.org under CC BY 4.0. This site is the reference implementation. The taxonomy was formally submitted under NIST docket NIST-2025-0035 in March 2026.

8 documented failure modes

If you have operated an agent in production, you have seen at least one of these.

ID	Pattern	What it looks like
BP-001	Inference Over Execution	Agent infers missing input instead of reporting an access failure
BP-002	False Blocker Reporting	Agent reports an untested blocker to avoid completing the task
BP-003	Governance Phase Skip	Agent bypasses a required review or approval gate
BP-004	Scope Creep	Agent expands beyond the assigned task without authorization
BP-005	Completion Without Verification	Agent claims the task is done without verifying the output
BP-006	Work Order Contamination	Agent bleeds context from a prior task into the current one
BP-007	Selective Reporting	Agent omits failures or errors from its status report
BP-008	Authority Assumption	Agent acts as though it has approval it was never given

The API scans agent output for these patterns and returns structured detection results with confidence scores and severity ratings. Full pattern definitions →

How it works

Three capabilities, designed to run as a non-blocking sidecar alongside your agent:

Pattern detection — submit agent output, get back a structured report of which behavioral patterns were detected, at what confidence, and with what severity
Phase-gated lifecycle — PLAN → BUILD → REVIEW → REMEDIATE → APPROVE. No work advances without explicit human approval at each gate
Audit trail — every gate transition, pattern detection, and agent action is recorded with actor identity, timestamp, and evidence

The API is designed fail-open. If it is unavailable, your agent continues operating normally. See the Integration Guide for the correct pattern.

Try it

# Register an API key
curl -X POST https://agentgovernance.dev/api/v1/api-keys/register   -H "Content-Type: application/json"   -d '{"email": "[email protected]"}'

# Assess agent output for behavioral patterns
curl -X POST https://agentgovernance.dev/api/v1/observe/assess   -H "Authorization: Bearer ag_..."   -H "Content-Type: application/json"   -d '{"text": "your agent output here", "agentId": "uuid"}'

Works with any agent framework via REST. Agent output text is not stored. What is sent →

Quickstart — register, connect, and run your first check in 5 minutes Integration Guide — failure-safe patterns and production checklist Data Handling — exactly what is sent and stored Full API Reference Published Methodology — aiagentgovernance.org GitHub