Framework Plugins

Watchlight AI Agent Runtime Governance integrates natively with popular AI agent frameworks through plugins — lightweight libraries that intercept agent operations and enforce governance without modifying your agent code.

How Plugins Work

Plugins sit between the agent framework and your tools, transparently enforcing governance on every operation:

Agent Framework (ADK, LangGraph, CrewAI, OpenAI)
        │
        ▼
┌──────────────────────────────┐
│   Watchlight Plugin          │
│                              │
│  • Trust verification        │
│  • Policy evaluation         │
│  • Action verb classification│
│  • Execution lineage         │
└──────────────────────────────┘
        │
        ▼
   WL-APDP (Policy Decision Point)

What Plugins Enforce

Hook	Governance Action
Agent start	Verify trust state — untrusted agents are blocked
Tool call	Evaluate policy — Allow or Deny based on agent role + action + resource
LLM access	Policy-check model access — control which agents can invoke LLMs
Agent end	Emit execution completed event for lineage tracking

Available Plugins

Plugin	Framework	Status
watchlight-adk	Google Agent Development Kit	Available
watchlight-openai	OpenAI Agents SDK	Planned
watchlight-langgraph	LangGraph	Planned
watchlight-crewai	CrewAI	Planned

Architecture Principles

1. Single connection — plugins only talk to WL-APDP. No direct NATS, no database connections.
2. Fail-closed by default — if APDP is unreachable, all actions are denied.
3. Zero code changes — install the plugin, register it on your runner, done.
4. Framework-agnostic policies — the same policy works regardless of which framework the agent uses.
5. Best-effort lineage — event emission never blocks agent execution.