architectural-skill

Define a LangGraph StateGraph supervisor with nodes for classify, plan, execute, evaluate, and deliver. The supervisor dispatches stateless agents from a registry, evaluates whether the goal is achieved, and replans up to five iterations before delivering results.

Wrap tool implementations as LangChain @tool functions, then convert them with Tool.from_langchain() before passing them to CrewAI crews. Each CrewAI crew has one agent with a role, goal, and backstory, and a TOOL_REGISTRY maps agent types to allowed tool subsets.

No. Agents must remain stateless, receive a message, and return a string without database access. Session and message history persistence is handled by the gateway's session manager using PostgreSQL with an in-memory fallback.

The gateway communicates with the orchestrator over HTTP using an httpx client, keeping services loosely coupled. The gateway exposes /message and /message/upload endpoints, while the orchestrator exposes /orchestrate on a separate port.

Replanning is capped at five iterations to prevent infinite loops when the evaluation node determines the goal is not achieved. Workflows requiring unbounded retries or long-running asynchronous tasks need additional scheduling logic outside the supervisor graph.