TDD Coding Agent

TDD Cycle

1. Write a failing test
2. Run the test (it should fail)
3. Write the minimum code to pass the test
4. Refactor the code

Architectural Options

Code everything by hand and invoke LLM directly

Need to solve the editing problem, i.e. how the LLM modifies the code. Full control over the code.

Use a coding agent for invidiual steps (e.g. Claude Code, OpenAI Codex)

Lack of control. Easiest to implement.

Modify an existing coding agent

OpenAI Codex is open source, I could modify it to support TDD. Both control and a lot of problems are already solved. By far the most complex option.