Prompt Improvement Agent - Better that Planning Agent Anti-Pattern

[wreuven]

There are recent clarifications that for many of the newer modern models, submitting a static plan is an anti-pattern that prevents the agent from dynamically improving the plan during execution.

I think there is still very much a role for a prompt-improvement agent but this improved prompt does not OUTPUT step-by-step planning. An interesting thing to note however is that the prompt-improvement agent presenting a theoretical plan could be a great part of the discussion towards improving the prompt.

For example consider this discussion.

agent - "I haven't decided exactly how this request should be carried out. I reserve the flexibility to change my plans on the fly as I get into it. However if I understand what you want, the steps will be something like ..."

user - "Based on your approach, I see I wasn't clear enough in my request... I need"

agent - "So would you agree that the prompt should be closer to ..."

user - "Yes. I also noticed in your plan that you consider web searching 5 sources as sufficient to research topic X and I would prefer that if you do research this topic, you would use 50 sources that include stack overflow and reddit.

[vedtam]

Yes plans can change, this is why I treat them as a living document—keeping it as a separate document that we both (me and the agent) improves over the lifetime of the execution.

Do you already have a draft for an agent file and a more concrete scenario?