Skip to main content

​
How framework wrapping works

In addition to its native create_agent(...) runtime, hexgate ships adapters that wrap agents built with OpenAI Agents SDK, LangChain / LangGraph, Google ADK, or Pydantic AI to add two things without touching the agent’s logic:
  1. Tool-call policy enforcement. Each tool the agent can invoke is gated by a PolicyEnforcer that returns a typed Decision (allow / deny / needs-approval) per call. Non-allow outcomes render as a [policy_denied] / [approval_required] marker the model sees as tool output (or, for pydantic_ai, a ModelRetry) rather than aborting the run, so the agent can recover.
  2. User-aware observability. Every run is traced through Langfuse with the active HexgateContext’s identity (user id, session id, role) propagated onto the spans.
Every wrapper also takes guards — before/after-tool functions that observe, rewrite arguments, or refuse a call — installed with guards=[...] (or HexgateRunner(guards=[...]) on the OpenAI and Google runners). The four integrations differ in shape because the underlying SDKs do: Role resolution happens at call time from the active HexgateContext contextvar — one wrapped agent serves many users concurrently because the scope is per-call. All adapters resolve the API key the same way: from the explicit api_key= argument, falling back to the HEXGATE_API_KEY environment variable.

​
Compatible framework versions

hexgate’s adapters are exercised against a framework-version matrix. Policy enforcement is verified across these ranges: Upper bounds are the latest versions tested, not hard ceilings — newer releases may work but aren’t verified yet. Minimums for pinned dependencies (openai-agents, langchain, langgraph, google-adk, pydantic-ai-slim) are enforced by hexgate’s requirements. langgraph>=1.1.10 is floored alongside langchain>=1.2.15: it’s the langgraph of that verified set, and older langgraph is the transitive import break that fails langchain 1.1.0/1.2.5.