So I've been reading a lot of KOLs like Karpathy make public statements about how they just write loops now and don't read code - in my experience frontier LLMs are not ready for unsupervised agentic architectures. The main reason is that agents tend to confabulate over long context horizons and misrepresent the verificatiojs they run on their claims.
This is especially damaging when the conclusions drawn by the agent are directionally load bearing, as in, leaning on the conclusion changes the direction of the project.
For instance, I was building a GTO poker solver a few weeks ago, and the solution could not close the exploitability gap (as in, no matter how many iterations I added, a true GTO solution could exploit gaps in my bots play and profit). Claude opus 4.8 kept asserting that this was a performance ceiling in the card abstraction I was using (grouping slightly different plays to the same clusters, a form of dimensionality reduction), and it cited academic literature to support its claims. I then increased the card abstraction and committed weeks more compute to solving the new setup, only to find out when the solve had finished that the bot had become even more exploitable (?!) only then did the agent recommend doing fanout debugging, and found a series of bugs in the game engine that was breaking the play.
A ground truth layer that gets called when the agent releases its task and audits the load bearing claims is highly effective in combating confabulation.
So I've been reading a lot of KOLs like Karpathy make public statements about how they just write loops now and don't read code - in my experience frontier LLMs are not ready for unsupervised agentic architectures. The main reason is that agents tend to confabulate over long context horizons and misrepresent the verificatiojs they run on their claims.
This is especially damaging when the conclusions drawn by the agent are directionally load bearing, as in, leaning on the conclusion changes the direction of the project.
For instance, I was building a GTO poker solver a few weeks ago, and the solution could not close the exploitability gap (as in, no matter how many iterations I added, a true GTO solution could exploit gaps in my bots play and profit). Claude opus 4.8 kept asserting that this was a performance ceiling in the card abstraction I was using (grouping slightly different plays to the same clusters, a form of dimensionality reduction), and it cited academic literature to support its claims. I then increased the card abstraction and committed weeks more compute to solving the new setup, only to find out when the solve had finished that the bot had become even more exploitable (?!) only then did the agent recommend doing fanout debugging, and found a series of bugs in the game engine that was breaking the play.
A ground truth layer that gets called when the agent releases its task and audits the load bearing claims is highly effective in combating confabulation.