What problem does it solve?
This Skill provides patterns and controls to prevent AI agents from performing unsafe or unauthorized actions by enforcing policies, filtering malicious input, rate limiting tool usage, and maintaining immutable audit trails for accountability and compliance.
Core Features & Use Cases
- Policy-driven tool control: Define composable, serializable policies to allow, deny, or flag tools and content patterns.
- Pre-flight intent classification: Detect data exfiltration, privilege escalation, prompt injection, and other dangerous intents before tool execution.
- Tool-level enforcement & audit: Wrap tool functions with decorators that enforce allowlists/blocklists, rate limits, content checks, human review gates, and append-only audit logging.
- Trust scoring for multi-agent workflows: Maintain per-agent trust scores with temporal decay to gate autonomous actions and delegate decisions.
- Framework integrations: Recipes for integrating governance with common agent frameworks (PydanticAI, CrewAI, OpenAI Agents) to secure production deployments.
Quick Start
Apply agent-governance to enforce a per-agent policy, run pre-flight intent checks on incoming prompts, and record all allowed and denied tool calls in an append-only audit trail.