2026-08-23 05:21 UTC
@norootcause@hachyderm.io I think your outlook makes a lot of sense. The premise of “if the agent is stuck it should page a human” also assumes the agent will somehow notice it is stuck. But given all I have seen so far in LLM Operators, they will not see a problem and take a step back, but happily jump to solve it. Even if, as you wrote, it doesn’t make sense or actually turns out to be harmful (“You are absolutely right, the core issue is different from what I first suspected”)
Replies (1)
-
@trenner@mastodon.social 2026-08-24 10:32
@bearbob@hachyderm.io @norootcause@hachyderm.io @saraislet@infosec.exchange Agreed. A core skill when coding with LLMs is knowing when to step in to tell the LLM to stop - they almost never find that limit correctly themselves. I have no idea why that would be any different for on-call automation. The technology has very limited ability to assess confidence in an approach by itself.