The issue isn’t really harness vs. no harness. IMO it’s about the lack of an internally generated sense of what to attend to. Yes, the KV cache accumulates state and its “attention” (if you can even call it that) changes with context. We’ve even managed to /kinda/ close the loop with agentic tool calling and ‘memory’ systems, but these just close the loop at the level of behavior rather than disposition. All agentic harnesses do is make an LLM responsive to the consequences of its actions without changing the tendencies by which it determines what to retain or avoid.
The ghost you can’t escape from at this point is the origin of that relevance. Where does the pull toward one thing mattering over another actually come from? If you ran Fable 5 on a Turing machine and rewound the tape to the exact same state with the exact same input (incl. PRNG seed), it would spit out the same output every time.
Everyone’s trying to outrun this problem by training more often or increasing model sizes. But all this does is inform your model, from the outside(!), what constitutes a better state. The thing that’s actually doing the determining remains unchanged. Congratulations, you’ve scaled the transition function and tape of your Turing machine until it requires every watt generated by ERCOT, and it still cannot, for the life of it, tell you why it should give a shit.
A trained model generating output from weights, a seed, and some context effectively has next-state that’s a total function of those three things. Whatever behavior appears as ‘selecting what is relevant’ is, underneath, just a transition rule executing, no matter how sophisticated or creative the output looks. It can be fully accounted for by what was fixed before it started executing. Which means whatever criterion it uses for determining what matters was inherited from a structure that was already in place before it encountered the situation.
No amount of pruning or post-training can fix this. These approaches just replace one externally supplied criterion with another. For a system to be truly adaptable, there would have to be some criterion by which it treats one possible change as preferable to another, and that criterion itself would have to come from... somewhere. You can even change your conception of ‘improvement’ (e.g. parameter count, harnesses, self-modification, hell, even its ability to spit out shitty best-selling romance novels onto Amazon) and you still haven’t explained where the normative distinction comes from. Every layer of this problem has its root in a preference that was supplied from somewhere else.
I genuinely don’t know if this issue bottoms out anywhere, at least for the way we currently build these systems. Perhaps the solution is still computable, maybe? Who knows what that would even look like. But I’m fairly confident that it isn’t a bigger tape. I hope nobody solves this in the near future because, well, I’d like to have a job...