Agent Governance Toolkit

Agent Governance Toolkit (AGT) adds a control layer around autonomous AI agents so that every tool call, message send, and delegation is checked in deterministic application code before it reaches the wire. It answers three operational questions: whether an action is allowed, which agent performed it, and how to prove afterwards what happened and under which policy.

The starting point is a govern() wrapper: you pass a tool function and a YAML policy, and each call is evaluated against the policy, written to an audit trail, and raised as GovernanceDenied when a rule blocks it. Rules can allow, deny, or require approval from named approvers. An AgentControl API and a Rust-backed Agent Control Specification runtime provide programmatic, stateless, fail-closed policy decisions for larger deployments.

The processing chain runs agent → policy engine (YAML, OPA, or Cedar) → identity (SPIFFE, DID, or mTLS) → tamper-evident audit log, and each layer is optional. Python carries the full package set; TypeScript, .NET, Rust, and Go SDKs implement the core of policy, identity, trust, and audit, and Copilot CLI, Claude Code, and OpenCode integrations are built on the TypeScript SDK.

Every major component is defined by an RFC 2119 specification with conformance tests, covering the policy engine, identity and trust, hypervisor execution control, SRE governance, the MCP security gateway, RL training governance, audit and compliance, and the wire protocol.

Features

• Policy enforcement: YAML rules with allow, deny, and require_approval actions, evaluated on every governed call
• Identity: per-agent identity through SPIFFE, DID, or mTLS so actions in multi-agent systems are attributable
• Audit log: tamper-evident decision records that capture the active policy, the request, and the verdict
• Execution sandboxing: the Agent Runtime package isolates execution across four privilege rings
• SRE controls: kill switch, SLO monitoring, chaos testing, and circuit breakers in the Agent SRE package
• MCP security gateway: tool poisoning detection, drift monitoring, typosquatting checks, and hidden-instruction scanning
• Compliance CLI: agt verify for OWASP checks, agt red-team scan for prompt-injection audits, and agt lint-policy for policy validation
• Shadow AI discovery: finds unregistered agents across processes, configs, and repositories
• Multi-language SDKs: Python (full stack), TypeScript, .NET, Rust, and Go, plus Claude Code, Copilot CLI, and OpenCode surfaces
• Specifications: RFC 2119 specs with 992 conformance tests and 29 architecture decision records