Every rule worth keeping came from something going wrong. The durable value of a retro isn't its narratives — it's its imperatives. If your post-mortem doesn't produce rules for next time, you shipped stories.
I specified a dataclass field name in a dispatch prompt. The agent built to spec, then stopped and flagged that the consuming interface expected a different name. The drift was on me, and it only took one grep to prevent.
Four agents writing code in the same git checkout. Ten stashes and 45 minutes of recovery later, the rule wasn't the lesson — the announcement that enforces it was.
An engineer agent dispatched to wire a module discovered the module didn't exist — only an empty __init__.py. The spec had merged two days earlier. Nobody had queued the build.
A rebuild's timeline is set by what you refuse to rebuild. Three days to ship a greenfield system worked because the cuts were in the requirements document before anyone felt the pressure to reverse them.
An agent caught a latent bug in legacy code the orchestrator's prompt didn't flag. That single act earned weight on their next flag — and that weighted flag caught two more bugs before they shipped. Trust compounds through a chain, not just a single delivery.
The bottleneck in AI-assisted development isn't writing code faster — it's thinking sequentially when the work isn't. Here's how dispatching three agents simultaneously collapsed three review cycles into one.
We shipped four bugs past code review, passing CI, and two AI reviewers in a single day. Here's what that taught me about the real limits of agentic coding — and the one discipline that would have caught all of them.
Google didn't release a universally smarter model. They released a more honest one. The specialist model era is here — and most engineering teams are still shopping wrong.