As far as I can tell, the only reason agents exist is because large context increase the probability of context poisoning, purely by the inability of these models to actually make conceptual decisions about the context.

I was interested in making a semi-automous skill improvement program for open code, and I wired up systemd to watch my skills directory; when a new skill appeared, it'd run a command prompt to improve it and cohere it to a skill specification.

It was told to make a lock file before making a skill, then remove the lock files. Multiple times it'd ignore that, make the skill, then lock and unlock on the same line. I also wanted to lock the skill from future improvements, but that context overode the skills locking, so instead I used the concept of marking the skills as readonly.

So in reality, agents only exist because of context poisoning and overlap; they're not some magicaly balm to improving the speed of work, or multiplying the effort, they simply prevent context poisoning from what's essentially subprocesses.

Once you realize that, you really have to scale back the reality because not only are they just dumb, they're not integrating any real information about what they're doing.