Path
Running and checking AI agents
Give an agent a real task and watch every step. Decide where it must stop. Report what it actually did.
- Prepares you for
- Agent supervision is the part of this market that is growing while simple labelling shrinks. It prepares you for agent-evaluation and workflow-operation work, and for the review side of safety projects.
- What you learn
- You give an agent a task with a real result. Then you do the part that pays. You watch the steps, catch the one that went wrong, and add the check that stops it happening again. At the end you hand the whole thing to somebody else with a note they can follow.
- Why this order
- You cannot check a run you have never watched, and you cannot hand over a run you cannot check. So it goes run, check, hand over.
- What an agent actually does16 minutes
- Write a task an agent can finish18 minutes
- Watch one run from end to end18 minutes
- Check the steps, not the ending20 minutes
- Decide where a run must stop20 minutes
- Count what a run costs18 minutes
- Add the check that stops the break22 minutes
- Hand a workflow over to somebody20 minutes
- Report a run in numbers22 minutes
Curriculum reviewed: