Implement a single-tool Python agent loop

Deliverableartifact

Python script (e.g., agent_single_tool.py) that implements a loop calling an LLM, parses a ReAct-style decision, invokes one safe tool function, maintains text state, and stops via final answer or max steps.

Goals

Run the script on a simple goal (e.g., small calculation or transformation) and observe at least one full Thought → Action → Observation cycle.
Show that the loop stops either when a final answer is produced or when a max step count is reached.
Demonstrate that the chosen tool is only called through the loop logic, not manually.
Print the evolving state/history to verify it records each step correctly.

Resources

🎬OpenAI Responses API: Add Custom Tools (Function Calling in Python)📄Function calling | OpenAI API 📰Implement a simple ReAct Agent using OpenAI function calling | Peter’s Notes 📰OpenAI Python API library 📰A simple Python implementation of the ReAct pattern for LLMs | Simon Willison’s TILs

Overview: what you’re building

You’re going to build agent_single_tool.py: a tiny ReAct-style Python agent that:

Talks to an LLM in a loop
Lets the LLM choose to call one safe tool (a calculator)
Keeps a text history of thoughts, actions, and observations
Stops when the model gives a final answer or hits a max number of steps
Logs everything so you can see what’s going on

This is the same core loop you see in ReAct examples like Peter Roelants’ simple ReAct agent and Simon Willison’s Python ReAct pattern, just trimmed down to a single safe tool.

We’ll do this in small pieces; you can paste everything into a single file and run it.

1 / 7