Start with one small task.
Before an agent takes a new job, give it a small, low risk task first. Read the result, then decide how much more to hand over.
A new job starts small
A new job is a guess about what the agent will do. A small task lets you test the guess on work you can look over in one sitting. You learn how the agent reads the job.
Pick a task with a clear check
Choose a task where you can tell by looking whether the result is right. Summarize one short document. Sort a few notes into folders. Draft one reply. If you cannot say what a good result looks like, pick a different task.
Write the three lists first
Before the agent starts, fill in May read, May change and Never touch for this one task. Keep them short. The permissions guide shows how to write them. No task is risk free, so the lists matter even when the work is small.
Read the result yourself
Open what the agent made and compare it with your check. Do not hand this part to another agent on the first run. You are learning what the job looks like when it goes well, and what it looks like when it does not. For later runs, see the review step.
Write down what you saw
Note what the agent did, what you expected, and any gap between them. A few lines is enough. Put them in the activity log so the next run starts from what you learned.
Fix the instructions, then run it again
If the result missed, change the job description, not the agent. Say the missing thing in plain words and run the same small task again. When the result matches your check, the job is ready for a slightly bigger task.
Check it: a made up trial run
- New job: tidy the notes folder.
- Small task: sort a few notes into folders by topic.
- May read: the notes folder. May change: which folder each note sits in. Never touch: what the notes say.
- Check: each note sits in the folder for its topic and nothing else changed.
- Result: most notes landed right. One sits in the wrong folder. Note it, add the topic rule to the instructions, run again.
One small task. One clear check. Then the next one.
