Agent 002 — Coding Agent

A small autonomous coding agent built from first principles in Python using the OpenAI Responses API.

What it can do

Agent 002 accepts a coding goal and autonomously decides which tools to use to complete it.

Its current tools are:

• list_files — inspect the project workspace
• read_file — read source code
• write_file — modify source code
• run_python — execute Python files and observe the result

Agent Loop

The core architecture is:

Goal → Decide → Act → Observe → Decide Again → Complete

For example:

User goal:

Find the bug, fix it, and verify that the fix works.

Agent 002 can autonomously perform:

1. List project files
2. Choose a relevant file
3. Read its source code
4. Identify a bug
5. Modify the code
6. Execute the program
7. Observe stdout, stderr, and the return code
8. Decide whether the task is complete

The sequence is not hard-coded. The model chooses tools based on the goal and observations it receives.

Safety Boundaries

Agent 002 includes several basic controls:

• File access restricted to the configured workspace
• Python-only execution
• 10-second execution timeout
• Maximum 10-step agent loop
• API key stored outside source code using .env

Example

Given a program containing:

return total / 0