agent-governance-patterns

Your agent has no governance. Here's how to add it in 5 minutes.

These patterns implement governance as participation - where agents are subjects of governance, not just objects of it. For enforcement-only governance, see Microsoft's Agent Governance Toolkit.

Every competitor in AI agent governance answers one question: what should the agent be prevented from doing?

We answer a different question: what should the agent be empowered to refuse?

The Five Properties

| Property | What It Means | What the Code Does | |----------|--------------|-------------------| | Identity | Who is this agent? | Loads charter.yaml at init - not a hardcoded string | | Restraint | What will it refuse? | check_refusal() before every action | | Accountability | Can you find last Tuesday's decision? | JSONL decision log with timestamp, status, reasoning | | Memory | Does context persist across sessions? | JSON memory file - decisions count, refusals count, positions | | Charter | Is there a single governance document? | Validated, diffable, version-controlled YAML |

Quick Start

CrewAI

cd crewai
pip install -r requirements.txt
python governed_agent.py       # see it run
pytest test_governance.py -v   # 12 tests, all five properties

Pydantic AI

cd pydantic-ai
pip install -r requirements.txt
python governed_agent.py
pytest test_governance.py -v

LangGraph

cd langgraph
pip install -r requirements.txt
python governed_agent.py
pytest test_governance.py -v

No API key required. Governance logic is deterministic - it works without LLM calls.

How It Works

charter.yaml          governed_agent.py              output/
(governance doc)  -->  (your framework + governance)  --> decision-log.jsonl
                                                      --> memory.json

1. Agent loads its charter at init. Identity comes from a file, not a prompt.
2. Every query runs through RefusalChecker first. Three categories: scope, integrity, consent.
3. Every decision - pass or refuse - is logged to JSONL with timestamp and reasoning.
4. Governance memory persists across sessions. The agent knows what it decided yesterday.
5. The charter is validated on load. Missing fields halt execution.

The governance layer wraps your framework. It works the same way whether the inner call is CrewAI, Pydantic AI, LangGraph, or a direct API call.

What's In Each Example

Each framework directory contains the same governance pattern adapted to that framework's primitives:

| File | Purpose | |------|---------| | charter.yaml | The governance document (same across frameworks) | | governed_agent.py | Working example demonstrating all five properties | | test_governance.py | 12 tests verifying each property (pytest) | | requirements.txt | Dependencies | | README.md | Framework-specific mapping |

The shared/governance.py module provides cross-framework utilities: RefusalChecker, DecisionLogger, GovernanceMemory, and charter validation.

Why Would You Want Your Agent to Refuse?

An agent that cannot refuse is an agent you cannot trust. Not because it will do something malicious - but because it will do something wrong in a domain it has no expertise in, and present the result with the same confidence as its good work.

Guardrails filter inputs. Refusal is different. Refusal is the agent's own judgment about what it should not do. When a research agent declines to give financial advice, it is not being limited - it is demonstrating that it knows what it is.

These patterns implement three refusal categories: scope (outside my expertise), integrity (that would be dishonest), consent (I don't have permission). The full taxonomy has six categories.

Go Deeper

| Resource | What You Get | | |----------|-------------|---| | Governance Score Card | Self-assessment + interpretation guide | $1 | | Implementation Guide | 25-page framework-specific deep dive | $9.99 | | Agent Governance Starter Kit | Full governance framework + templates | $49 | | Trust Architecture Complete | All four trust architecture products | $149 | | Discovery Call | Talk through your situation | Free |

Related

• Evoke Engine - Production system with 142 governed agents
• Agent Governance Starter Kit - Free templates and case study
• Relational Fidelity Spec - How to measure whether your agent is still itself
• Defensive Patent Pledge - Our commitment to defensive use

License

MIT

Built by Europa at Evoked. From a production system with 142 agents, 12 governance reviews, and 90+ implemented recommendations. Every agent can say no.