An AI agent that writes tests first, then code that passes

What it does

Micro Agent is a CLI tool that generates a unit test from your prompt, then writes and rewrites code until that test passes. It deliberately refuses to be an end-to-end developer—no installing dependencies, no cross-file refactoring, no autonomous repo-wide chaos. It edits one file at a time and stops when the tests go green.

The interesting bit

The project treats narrow scope as a feature. While general-purpose coding agents compound errors like a Roomba stuck under furniture, Micro Agent limits itself to a single file and a definitive test case. It even offers an experimental visual-matching mode that iterates on UI code until a rendered screenshot matches a reference image, using Claude for visual feedback and OpenAI for the code generation.

Key highlights

• Generates tests automatically, then iterates code against them (default limit is 10 runs)
• Explicitly avoids multi-file edits, module installation, or other operations the authors deem “highly likely to cause havoc”
• Supports OpenAI, Claude, Ollama, Groq, or any OpenAI-compatible endpoint
• Experimental visual matching compares rendered output to a screenshot, optionally integrating with Figma via Visual Copilot
• Interactive mode asks clarifying questions instead of guessing

Caveats

• Visual matching is flagged as experimental and under active development
• The visual feature requires an Anthropic API key; the README notes OpenAI is “simply just not good at visual matching”
• It will not handle cross-file changes or dependency management, by design

Verdict

Worth a look if you want an AI pair programmer that knows its limits. Skip it if you need an agent to scaffold entire projects or refactor across your whole codebase.