Fix a LangGraph-based AI agent (tools, state, retries)

We have an AI agent built with LangGraph, Python and OpenAI APIs. The core is done, but parts of the workflow are unreliable: tool execution, state handling, API integrations and response quality.

You will review the architecture, find the root causes, fix node, edge and routing bugs, repair tool and function calling, improve error handling and retries, test the whole workflow and leave clean, maintainable code.