Agent Governance Toolkit

Agent Governance Toolkit (AGT) enforces policy, identity, sandboxing, and reliability controls on autonomous AI agents in deterministic application code rather than in the prompt. Every tool call, message send, and delegation is intercepted before it leaves the agent, so an action the policy denies never executes. It addresses three questions: whether an action is allowed, which agent took it, and whether that can be proven to an auditor afterward.

Wrapping a tool function with the governance call evaluates a YAML policy on every call, writes the decision to an audit trail, and raises a denial when the policy blocks the action. The pipeline runs agent requests through a policy engine (YAML, OPA, or Cedar rules), an identity layer (SPIFFE, DIDs, and mTLS), and a tamper-evident audit log. Every layer is optional. Policy evaluation is delegated to the Agent Control Specification, a stateless, fail-closed decision runtime with a Rust core, and AGT keeps the host APIs and framework adapters around it.

SDKs exist for Python (the full stack), TypeScript, .NET, Rust, and Go, plus governance plugins for coding agents (Claude Code, Codex CLI, OpenCode, and GitHub Copilot CLI). Framework integrations include native middleware for Microsoft Agent Framework and Semantic Kernel and adapters or middleware for AutoGen, LangGraph and LangChain, CrewAI, OpenAI Agents SDK, Agent Development Kit (ADK), LlamaIndex, Haystack, Mastra, and Dify. Enforcement runs in the same process as the agent, so the project recommends a separate container per agent for OS-level isolation.

Each major component has a formal RFC 2119 specification backed by 992 conformance tests, and the controls are mapped to the OWASP Agentic AI Top 10, NIST AI RMF, the EU AI Act, and SOC 2. The project is in public preview and may have breaking changes before general availability.