Tutorial
Eval-driven development for coding agents
How to build an evaluation harness before you build the agent, so you can tell whether a change made things better or just different.
Most agent work fails not because the model is weak but because nobody can tell whether a change helped. This tutorial builds a small eval harness: a fixed set of repository tasks, a deterministic grader, and a scoreboard. Then it shows how to use that harness to make a prompt change you can actually defend.
evals
testing
agents
Related
Writing your first MCP server
A step-by-step guide to building, testing and shipping a Model Context Protocol server that exposes your own tools to an agent.
Prompting patterns for large codebase edits
Context selection, chunking and verification strategies for making an agent safe to point at a million-line repository.