Agents should know about their limitations while creating new agents.

[kazunator]

I find it that LLMs even with updated knowledge cutoff don't really have a good grasp over their own limitations. One of which is the context range on which they perform best. To my understanding, chatgpt4 works best when performing tasks that do not go beyond the 200-300 token range. Anything beyond that and it will try to rush the task and finish it even when not complete. This limitation should be properly communicated to the first layer agents that will create agents below them. And these new agents should also keep these limitations in mind when creating new agents.
Of course, the obvious limitations shouldn't be communicated as that will only create fluff and noise. Only communicate stuff that it doesn't know implicitely.

[daveshap]

Yes this should be covered in the standard docs given to all agents. It's call an AGENT MODEL

[kazunator]

Forgive me I'm confused. Are you saying this should be a doc called AGENT MODEL? or is there already a doc called AGENT MODEL? Looked around and couldn't find it

[daveshap]

Ask the concierge

[kazunator]

https://chat.openai.com/share/411af6aa-8837-451f-adfd-6b36c6a3addb our discussion