General purpose agentic loop in 40 lines of Python (softwaredoug.com)

A short metaprogramming recipe shows you can build a general-purpose agentic loop in ~40 lines of Python that treats ordinary annotated Python functions as LLM-callable “tools” without the usual manual plumbing. The core is a make_tool_adapter that uses introspection and type annotations (Pydantic models) to auto-generate a tool_spec (name from the function, docstring as description, JSON-schema-like parameters including Literal enums) and a wrapper that deserializes LLM-provided arguments, calls the function, and returns both the Python result and JSON output. The example wires these adapters into a simple agentic_loop that sends system/input to the OpenAI Responses API, detects function_call outputs, validates arguments via ArgsModel.model_validate_json, invokes the wrapped function, appends a function_call_output entry (model_dump_json), and repeats until no more tool calls appear. This approach is significant because it removes repetitive packaging/unpacking work and lets you prototype flexible agent workflows without a heavy framework or MCP server. Key technical implications: it relies on robust type annotations and Pydantic for safe validation/serialization, maps Python types -> tool specs for the LLM, and supports easy extension (parallelism, async) but isn’t intended as a production-grade orchestration system. It’s a practical, minimal pattern for rapid experimentation with agentic tool calling and function-level integration with modern LLM tool-call APIs.