Phase 3: Build

Input: docs/agent-design.md. Output: working code, and traces you can read.

Agent code fails differently from ordinary code. It usually runs, returns something plausible, and is wrong in ways no exception surfaces. So the build phase has one extra obligation: make the agent's behaviour observable from the first commit, because you cannot debug what you cannot see, and printing the final answer tells you almost nothing.

1. Build the tools before the loop

Tools are ordinary code and should be tested as ordinary code. Get them correct and fast in isolation — direct unit tests, no model involved.