Code review agent

The brief

A team's senior developers spend hours on routine review comments. Build an agent that reads a pull request, runs the tests and linters through tools, and drafts review comments, which a human approves before anything is posted. Good looks like: an agent with clear tools and permissions, an approval step, a test set of real pull requests, and its success rate and cost per review.

What you will build

1. Design the agent's tools: read diff, run tests, run linter, post comment
2. Build the agent with LangGraph, with state and a human approval step
3. Expose the repository tools through an MCP server with limited permissions
4. Build a test set of pull requests with known issues
5. Measure issues found, false comments and cost per review
6. Guard against prompt injection hidden in code or comments

Free, open data to build it with

• SWE-bench ↗ Hugging Face (Princeton NLP)

Check each dataset's licence before you publish your work. Your mentor gives your team its own version of the brief (data, scope or region), so every team's project is its own.