2 ms·
It doesn't have to. But usually verifying the answer is cheaper than finding the answer, so it can be done. The prompt given to the sub-agent can be a question
by anshorei 20d ago
It doesn't have to.
But usually verifying the answer is cheaper than finding the answer, so it can be done. The prompt given to the sub-agent can be a question with "and explain your answer", the answer to which will still consume far less context than figuring out the answer.
- jurgenburgen 19d ago> It doesn't have to. Of course it has to. The subagent might have misunderstood and spent 250k tokens smoking crack. If that output is blindly trusted the parent agent might go on to burn millions of tokens going in the wrong direction, completely negating the point of using a subagent in the first place.