The idea
You can trust a team of agents when you can stop watching them and still know what’s true.
That doesn’t come from a smarter model. It comes from the way the work is set up around the model:
- “Done” needs evidence, not a claim.
- Status comes from records, not from what the agent says.
- Agents settle what your instructions already answer and bring you only real decisions.
- What an agent may do is written down, not assumed.
The talk follows two tools I built. Case taught me to trust one agent’s pull request. Riker is a teammate that runs other agents for me and comes back with receipts and questions. You don’t need either one for these lessons. Each lesson is a habit you can start with Claude Code or Codex today.
How to use these lessons
Each lesson is short and ends with an exercise on a real task of yours. Pick one small task now that you’d normally hand to an agent and then watch. Use it in every lesson.
- Write the definition of done first
- Ask for evidence, not a summary
- Use a second session as the reviewer
- Write down what it may decide alone
- Design your interrupts
Exercise: notice where you watch
Next time you hand a task to an agent, write down every time you check on it. Next to each one, write what you were worried about. Those notes tell you which lesson to start with.