Deciding and doing

The policy engine

Policy as code

Context-aware policy

The audit trail

Learn Microsoft's Agent Governance Toolkit from scratch: intercept every tool call, write allow-lists and YAML policy, record a tamper-evident audit trail, and gate it all in CI. Every lesson runs with no API key.

Agent Governance Toolkit

Where the course goes

Agent Governance Toolkit tutorial for beginners

What you are going to build

Why an agent needs governance · Agent Governance Toolkit

An agent that just does what it decides

A tool-calling agent with no model · Agent Governance Toolkit

The agent that decides

Install the Agent Governance Toolkit

Installing the toolkit, and the name it imports under

What one call passes through

Intercept a tool call · Agent Governance Toolkit

The choke point

Default deny and allow-lists · Agent Governance Toolkit

Default deny, and the allow-list

Argument-level rules · Agent Governance Toolkit

Checking the arguments, not just the tool

Interception versus validation · Agent Governance Toolkit

Two paths through one engine

The mistake this lesson exists to prevent

Custom policy rules · Agent Governance Toolkit

Writing a rule of your own

Freezing a policy engine · Agent Governance Toolkit

Stopping an agent from editing its own policy

Declarative YAML policy · Agent Governance Toolkit

A policy in a file

Policy conditions and operators · Agent Governance Toolkit

Conditions, and the field that is not there

Rule priority and ordering · Agent Governance Toolkit

Which rule wins

Policy defaults and actions · Agent Governance Toolkit

Defaults, and the four actions

Policy scope and its silent no-op · Agent Governance Toolkit

Scope, and the rule that applies to everything

Conditional permissions · Agent Governance Toolkit

Allowed only when the context agrees

Argument conditions in policy · Agent Governance Toolkit

Conditions on the arguments themselves

Updating policy context at runtime · Agent Governance Toolkit

Context that changes while it runs

Risk-based policy · Agent Governance Toolkit

Scoring a call instead of naming it

The empty audit log · Agent Governance Toolkit

Where the audit trail went

Recording every decision · Agent Governance Toolkit

The flight recorder

Querying the audit log · Agent Governance Toolkit

Buffering, flushing and asking questions

Tamper-evident audit logs · Agent Governance Toolkit

Proving the log was not edited

Deciding and doing quiz

Your loop calls <code>intercept_tool_execution</code> and ignores what it returns. What happens to a call the policy refuses?