← Back to context

Comment by jes5199

5 days ago

there’s a sweet spot in there, it’s not “as narrow as possible” - the most productive thing is to assign the largest possible tasks that are just short of the limit where the agents become stupid. this is hard to hit, and a moving target!

Exactly. When you get a new tool or a new model, ask it for things the previous one failed at until you find the new ceiling.