I had an interesting experience today: Claude Code asked me a question that required consideration and, after I didn't reply within a short time frame, it said "You're busy. I'll just assume..." and simply went on.

Even though I have regularly run with "Accept Edits" or even occasionally "Auto Mode" (once a plan was agreed), I consider this behaviour to be dramatically different. The planning and discussion phase of working with an LLM seems to me the most important: you build the context in which code will be written. The decisions and observations made in this phase are the work. The better they are, the better the output and long term outcomes. And it feeds through into future planning and future decisions. It's either a virtuous or a vicious cycle, depending on the level of thought given to each decision. I will often go for a walk (or a mere piss) to think over a specific question. Often the question itself causes me to reverse course or unmakes a prior assumption.

"You're busy, I'll just assume..." means I can no longer be sure in a long conversation whether every decision was mine or not. I understand to some people they may not see a difference between this and accept edits, but to my mind what I work on is decisions - the actual code written may or may not match the decisions, but that's something I will work through in documentation and testing. I see the logic of this from Antropic's point of view: it moves the agent towards autonomy without going full Claw.

I think though that this step is going to prove a lot less profitable. These decisions are not of the same order as code design and my experience is that the more critical the decision, the wronger the Agent tends to get it.

That's not how it works; when Claude prompts you for an answer, it's not going to wake itself up if you fail to answer in time unless you have it on a cron.

[flagged]