Interesting idea, Sara! Curious, how would the agent incorporate the feedback? I’m guessing it could be added to the prompt (but context length could be an issue) or training a new model (which might be too resource intensive)? I could be missing something obvious.
I recently wrote about quite a different idea, where agents should not show their steps because it overwhelms the user, especially when there are too many steps. And ideally, we shouldn’t need to monitor our agents (thought reliability is still an issue for now). https://letters.alfredlua.com/p/alfred-intelligence-6-ai-agents-should
A lot of people are working on agent memory, including OpenAI themselves w the Operator API launch soon... I don't have a good POV on what the best method is, time will tell
As a person interested in AI agents and UX design, this writing was very enlightening. Thanks for sharing your work
Interesting idea, Sara! Curious, how would the agent incorporate the feedback? I’m guessing it could be added to the prompt (but context length could be an issue) or training a new model (which might be too resource intensive)? I could be missing something obvious.
I recently wrote about quite a different idea, where agents should not show their steps because it overwhelms the user, especially when there are too many steps. And ideally, we shouldn’t need to monitor our agents (thought reliability is still an issue for now). https://letters.alfredlua.com/p/alfred-intelligence-6-ai-agents-should
A lot of people are working on agent memory, including OpenAI themselves w the Operator API launch soon... I don't have a good POV on what the best method is, time will tell