microsoft

agent-governance-toolkit: Govern AI Agent Actions

Microsoft's Python toolkit uses the upstream Agent Control Specification policy runtime to govern autonomous AI agents. It adds policy enforcement, identity, audit, sandboxing, and reliability controls, with integrations for multiple agent frameworks and language SDKs.

View on GitHub Compare
Stars
6.6k
Forks
1.1k
Stars, last 30 days
+131
Last commit
today
Contributors
100+

Our take

Activeregular commits and releases

Health score

68/100

How it is scored
Activity
79
Community
86
Issues
21
Pull requests
84

+131 stars in the last 30 days

  • Merges more pull requests than 95% of the projects we track
  • More contributors than 78% of the projects we track

AGT is an actively maintained, broad governance toolkit with substantial cross-language and framework coverage, but its public-preview status and application-layer security boundary deserve careful evaluation.

Good fit if

  • You need deterministic policy checks and audit records around autonomous agents' tool actions.
  • You want governance integrations across frameworks, with Python as the primary full-stack option.
  • Your team can evaluate a public-preview project and deploy agents in separate containers when OS-level isolation is needed.

Look elsewhere if

  • You need governance enforced at an OS or process boundary by this toolkit alone. The README says AGT's middleware and agents share a process boundary.
  • You need a stable GA release with no expected breaking changes.
  • You need the full feature set in a non-Python SDK. The README says Python has the full stack, while other language SDKs cover core governance.
All health signals
Last commit2026-11-10 (today)
Commits, last 90 days100+
Releases, last 12 months19 (latest v4.1.0, 2026-06-09)
Contributors100+ (top contributor: 35% of commits)
Issues closed, last 90 days0+
Pull requests merged, last 90 days93 (typically merged in 0 days)
Project age8 months

Checked on 2026-11-11 with the GitHub API.

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

6,566

Forks

1,127

Change since Nov 15, 2026: -1 stars·0 forks

Overview

Microsoft's Agent Governance Toolkit (AGT) is a cross-framework toolkit for governing autonomous AI agents. Its policy layer uses the upstream Agent Control Specification (ACS) runtime, while the toolkit adds identity, audit, execution sandboxing, and reliability controls. It aims to control tool calls and other agent actions in application code, rather than relying on prompts to keep an agent within bounds.

The project is intended for teams deploying agents that can call tools, access services, or delegate work, especially where they need to identify agents and retain records of governance decisions. Python has the broadest feature set, and the repository also provides SDKs and integrations for TypeScript, .NET, Rust, Go, and several agent frameworks. AGT is in public preview, so breaking changes remain possible.

Key Features

  • Evaluate tool actions against policies, with allow, deny, and approval outcomes.
  • Record governance decisions in an audit trail.
  • Provide zero-trust agent identity and support trust and delegation controls.
  • Add execution controls, including privilege rings, command denylisting, and termination controls.
  • Connect governance to frameworks including Microsoft Agent Framework, Semantic Kernel, AutoGen, LangGraph, CrewAI, and the OpenAI Agents SDK.
  • Include an MCP security gateway for detecting issues such as tool poisoning and drift.
  • Provide CLI commands for installation checks, policy linting, compliance verification, and prompt-injection audits.

Use Cases

  • A platform team can require approval for an agent's email or other consequential tool calls while allowing routine reads.
  • Security engineers can identify which agent requested an action and review its recorded policy decision during incident response.
  • Teams using multiple agent frameworks can apply governance controls through framework adapters rather than building separate controls for each integration.
  • Compliance teams can use the verification tools and mappings to assemble evidence against standards such as OWASP Agentic AI Top 10, NIST AI RMF, and SOC 2.

What you need

Detected in the repository

  • Python (main language on GitHub)
  • A Dockerfile, so it can run in a container; a Compose file sets up the related services
  • A test suite and automated checks on GitHub Actions

License in plain words

MITpermissive

  • Commercial use: yes
  • Modify and redistribute: yes
  • You must keep: the copyright and license notice
  • Share your changes: no

A summary, not legal advice: the LICENSE file is what applies.

Getting Started

Python 3.11 or later is recommended for the quick start:

pip install "agent-governance-toolkit[full]"

See the README and quickstart guide for setup and framework-specific examples.

Alternatives

  • bernstein: Bernstein governs workflows with a deterministic scheduler and replayable run history, rather than extending the upstream Agent Control Specification runtime.
  • OpenBot: OpenBot is a self-hosted agent workspace whose policy gateway gates actions, while agent-governance-toolkit provides governance controls for integrating into agent frameworks.
ProjectLanguageLicenseStarsStatus
agent-governance-toolkitPythonMIT6.6kActive
bernsteinPythonApache-2.01.2kActive
OpenBotTypeScriptMIT6.8kNew

Considerations

  • The project is marked public preview, and its documentation warns that breaking changes may occur before general availability.
  • Governance is enforced in application middleware, not at the OS kernel level. The README recommends running each agent in a separate container for OS-level isolation.
  • The full Python quick-start uses the [full] extra; the base package provides the compliance CLI, while governance modules are in the core distribution.
  • The repository reports active development, with frequent recent commits and releases and a contributor base not dominated by one person. That activity is encouraging, but does not remove the preview-stage compatibility risk.
  • The project is MIT licensed. Review the license and third-party framework or service data-handling practices as part of deployment planning.

Found this useful?

Share it with someone who would like agent-governance-toolkit.

Comparisons