Developing an AI Agent (www.dolthub.com)

A developer set out to build an AI agent that can import and work with local data, documenting the trials and a working approach using the OpenAI Python Agent SDK. They show how simple it is to spin up an Agent and Runner, but quickly demonstrate a common pitfall: LLMs are stateless, so follow-up prompts lose context unless you provide a session. By switching to an SDK Session (e.g., SQLiteSession) the agent preserves conversation history and can correctly do follow-ups like “add 10 to the number you gave me.” The post underscores that agent frameworks simplify LLM requests but hide crucial plumbing — knowing the underlying API improves debugging and design choices. Technically, the write-up highlights function-calling tools (@function_tool) to expose file I/O (a read_file example reads CSV bytes), explains how the SDK encodes function signatures/docstrings into the model request, and how the model can request function invocation; the SDK runs the function, injects the output back into context, and re-invokes the model. The author warns this loop can rapidly exhaust an LLM’s context window when importing many files, and demonstrates session and memory abstractions (Session interface with get_items/add_items/clear, a MemorySession example) plus a simple filter that prunes old function calls/responses. Key takeaways for ML/AI practitioners: manage state explicitly, design tool interfaces carefully, and implement context-pruning/summarization strategies to scale data-import agents.