LLMs are improving at enormous rates with potential leaks of self improvements getting out of hand and mismanaged by the frontier labs. However, as I said before, documentation will always be the bane of LLMs or really even any team communication.

For example, I was asking Claude Code to scrape some content. In the middle, I asked for a progress update, and it found a performance issue in SQLite, so it added an index. With 7 hours left from 3 days of the scrape, estimated by Claude Code, I asked for another progress update, and it found another small progress improvement. I’m not in any rush, but it decided to make another update. I have plenty of time to let it finish, especially with 7 hours left. I wanted to reduce the risk of any potential mistakes it introduces, especially with such short amount of time left. But Claude isn’t calculating the risk; sometimes, I do notice it does. But then it also doesn’t know the time context.

The thing is, if I explicitly told Claude there is plenty of time, then it’s fine. I think if someone copilots their AI agents closely, because, nowadays, at least for me, I’ve become more reliant on its autonomy being mostly perfect, then a user can input feedback that I’ve described above.

However, my point isn’t that LLMs can’t calculate risk and time context. It’s all still up to documentation and context, just as team communication requires.

The upper limits of LLMs (for digital tasks at least) may now just be humans and their context itself.