PythonOpenAIFunction CallingAgent LoopTool Use

Mini Agents — Tool-Use Loop

OpenAI function-calling agent that plans and solves problems using a live todo list

Overview

Built a minimal but complete AI agent using OpenAI's function-calling API. The agent receives an open-ended problem, autonomously breaks it into steps via a todo list, executes each step in order, and returns a final answer — all without human intervention. The project demonstrates the core loop behind every tool-using agent: model decides → tool executes → result feeds back → repeat until done.

Live Demo

Mini Agents — Live Demo

Type any problem. The agent plans steps, executes them, and returns an answer.

Try:

Results will appear here

Methods

Defined two Python tools: create_todos (adds a list of steps) and mark_complete (marks a step done with notes)

Described each tool as a JSON schema and registered them with the OpenAI chat completions API

Implemented dynamic dispatch via globals() — tool name from the model maps directly to a Python function call

Built the agent loop: call model → if finish_reason is tool_calls, execute tools and append results → repeat until text response

Used Rich console markup for colored, strikethrough-rendered todo output

Tested on a multi-step math problem (two-train meeting time) requiring estimation and sequential reasoning

Key Findings

The model reliably decomposes open-ended problems into ordered, executable steps before acting
Dynamic dispatch via globals() keeps the loop generic — any registered Python function becomes a callable tool
The agent self-corrects by marking steps complete with notes, creating a readable audit trail of its reasoning
Structured system prompting (plan first, then execute, no clarifying questions) is key to autonomous behavior

Results

Full agent loop in under 60 lines of Python — no framework required

Agent correctly solved the two-train problem end-to-end: estimated Boston–NYC distance, computed meeting time

Todo list rendered live with Rich — completed steps shown in green strikethrough as the agent works

Reusable pattern: swap tools and system prompt to build any tool-using agent on the same loop

Ask About This Project

Have a technical question? Ask here.

← All projectsEric Jin