Post #3636687

Agentic AI has "guardrails" (e.g. you have to explicitly say "don't delete all the files on my computer" if you don't want it to delete all your files). One of these guardrails is asking users to confirm whether to do things. Claude Code apparently has now decided that if you take over sixty seconds to answer a question, it'll just go ahead. Turns out guardrails get in the way of consuming tokens, and you must consume tokens in order for the business model to work. https://github.com/anthropics/claude-code/issues/73125