A hundred agents in, the hard parts were rarely the clever parts. They were naming, scoping and deciding what done means.
Define done before you define the agent
The agents that struggled were the ones without a clear finish line. If a person cannot say what a finished job looks like, an agent cannot either.
Start narrow, then widen the scope
Every successful rollout started with one team, one job and tight permissions. Scope grew only after a month of clean traces.
The checklist we start with
- Write down what a finished run produces
- List every tool the agent may touch, and nothing else
- Set a budget per run before the first test
- Decide which actions need a person
- Replay the first fifty runs by hand
The best agents we shipped were the most boring ones to watch. They did one job, the same way, every time.
- Lessons
- Production