<p>The deeper I dive into the development of AI agents, the more questions arise, as well as the understanding that it’s far from being as simple and easy as it seemed at the beginning.</p>
<p>In general, an AI agent can be viewed as just a worker in an office who has a profession, a set of responsibilities, job descriptions, and access rights.</p>
<p>The question I’m currently trying to figure out is how autonomous the agents should be. Should they operate in the Apple way (where you are given only what you need and nothing more, just like in Apple Corporation), or should the agent have access to distributed memory where both the correspondence with the user and the results of previous agents’ work are stored?</p>
<p>On one hand, I want to create a super focus, as it’s clear that the less context used, the higher the quality. But then you have to rely on the results of previous agents (who might mess things up). If you don’t trust anyone (any reworking or filtering of information only leads to a decrease in signal) - there’s a big chance to lower quality by cramming too much information into one agent. Most likely, the answer lies in balance.</p>
<p>There are surely some scientific papers with A/B tests, articles, etc. Have you encountered similar thoughts? What did you decide, what do you think? Share your experiences. I believe the comments on this post will be very helpful to many.</p>
<p>#ai #agi #agent</p>
<p><a href="https://t.me/+OvImEUmA7W5mYTRi">————————— Мысли Рвачева —————————</a></p>