About this project

Agent Governance Toolkit (AGT) is a Microsoft open-source project for governing autonomous AI agents in production. Its premise is that prompt-level safety instructions are not a reliable control surface, so governance is enforced deterministically in application code before an agent's intent reaches a tool, message channel or another agent. Core capabilities - Policy enforcement: YAML policies (with OPA/Cedar mentioned) evaluated on every tool call; decisions can allow, deny or require human approval. A two-line `govern()` wrapper checks, logs and raises `GovernanceDenied` on blocked actions. - Identity and trust: zero-trust identity using SPIFFE/DID/mTLS concepts, agent discovery, routing and trust scoring across a mesh, plus delegation chains. - Execution sandboxing: an Agent Runtime with four privilege rings, saga orchestration, termination control and command denylist enforcement. - Reliability engineering (Agent SRE): kill switch, SLO monitoring, error budgets, chaos testing and circuit breakers. - Audit and compliance: tamper-evident (Merkle) audit trails, decision records, OWASP verification, policy linting and integrity checks via an `agt` CLI. - MCP security gateway: tool poisoning detection, drift monitoring, typosquatting and hidden-instruction scanning. - Additional tooling: shadow AI discovery, a Streamlit governance dashboard, a 12-vector prompt injection evaluator, and a contributor reputation GitHub Action. Packages and languages Python is the full stack, distributed as consolidated packages (core, runtime, sre, cli, and a `[full]` meta-package). SDKs also exist for TypeScript, .NET (including an MCP extension), Rust and Go, covering core policy, identity, trust and audit. First-party developer surfaces include plugins/installers for Claude Code, GitHub Copilot CLI, Codex CLI and OpenCode. Framework integrations Native middleware for Microsoft Agent Framework and Semantic Kernel; adapters or middleware for AutoGen, LangGraph/LangChain, CrewAI, OpenAI Agents SDK, Google ADK, LlamaIndex, Haystack, Mastra and Dify; deployment guidance for Azure AI Foundry. Specifications and standards Ten formal RFC 2119 specifications with a stated 992 conformance tests and 29 architecture decision records. Compliance mappings are provided for OWASP Agentic AI Top 10, NIST AI RMF 1.0, EU AI Act, SOC 2, AARM Extended and ATF. Security posture and limitations The README states governance runs at the application middleware layer, not the OS kernel, so the policy engine and agents share a process boundary; it recommends running each agent in a separate container for OS-level isolation. It also links a Known Limitations document. The project is labeled a public preview and may introduce breaking changes before general availability. Tooling includes CodeQL, Gitleaks, ClusterFuzzLite fuzz targets, Dependabot and OpenSSF Scorecard.