Under the Hood of a Coding Agent: The Loop, Tools and Reasoning

Takeaways

• Models have been trained to use tools, and tools such as read file, write file, list directory and run tests are what make a coding harness possible.
• In a demo on the OpenAI developer platform he defines those tools and asks a model to fix security problems in an app file; the model's own visible reasoning decides which tool to call next.
• The loop: we ask, the model reasons and requests a tool, we run it and return the result, and it repeats until the model finishes.
• It is not deterministic: on one run the model asked for the requirements file, and on another it did not.
• He then rebuilds the same review agent in Python with the LangChain agents framework and a sandbox of tools, and the file is patched.