agent-governance

<!-- BSV Скил : agent-governance TL;DR : Добавляет governance-слой к AI-агентам: политики, intent-классификация, аудит, trust-скоры Вызов : /agent-governance, agent safety, tool access control НЕ для : output guardrails (это pre-execution safety), UI/UX-паттерны, LLM fine-tuning --> # Agent Governance Patterns Patterns and techniques for adding governance, safety, and trust controls to AI agent systems. Works with any agent framework: PydanticAI, CrewAI, OpenAI Agents SDK, LangChain, AutoGen. ## Core Design Principle > Intent classification happens **before** tool execution — a pre-flight safety check. This is fundamentally different from output guardrails which only check **after** generation. --- ## When to Use This Skill - Building agents that call external tools (APIs, databases, file systems) - Implementing policy-based access controls for agent tool usage - Detecting dangerous or adversarial prompts before execution - Creating trust scoring for multi-agent pipelines - Building compliance-grade audit trails - Enforcing rate limits, content filters, or tool restrictions --- ## Pattern 1 — Governance Policy (Declarative) Define allowed/blocked tools and content filters in YAML or as a Python dataclass: ```yaml # governance_policy.yaml governance: level: standard # open | standard | strict | locked allowed_tools: - web_search - read_file - write_file blocked_tools: - execute_shell - delete_database content_filters: - p...