Day 2 of 5: the code
Day 2 of 5. The code.
The agent that plans the work never writes the code. Of all the rules we've added this year, that one changed more for us than any model upgrade.
Rune, our lead agent, writes the spec, the plan and the checks. The build goes to separate builder agents, and each builder gets its own git worktree. No two builders ever write to the same checkout. We learned that one the hard way.
The other rules came the same way. If a new idea turns up halfway through a build, it goes to a new builder and the running one is left alone to finish. A bug fix starts with a test that fails, and if we can't make it fail we haven't found the bug yet. The model is picked by how hard the job is, so small mechanical work goes to a cheap model and the top model only sees the hardest items, after the spec and plan have passed review. Urgency doesn't come into it. And a builder's progress lives in files, so any session can die and the next one picks up where it stopped.
A builder isn't finished when it compiles. It loops with a reviewer until the checks pass, opens a PR, and then someone else judges it.
Tomorrow, review, and why the reviewer is never the model that wrote the code.
If you've got more than one agent writing code in the same repo, how are you keeping them off each other's changes?
