Agentic coding tools — Claude Code on the web, Cursor's agents, long tool-using runs in any chat — introduced a new kind of dead time. The agent works unattended for twenty minutes, then silently stops: it wants permission to run a command, it asked which of two approaches you prefer, or it simply finished and is waiting. If you notice in ten seconds, the run flows. If you notice in ten minutes, you've paid the agent tax.
The states that actually need you
- Needs permission. The agent is asking to run a command or edit outside its sandbox. Every minute unanswered is a minute of paid-for autonomy wasted.
- Asked a question. A genuine fork — “option A or B?” — that no amount of waiting resolves.
- Finished with undone work. The reply reads as done but ends with “remaining steps…”. Technically finished, actually paused.
- Waiting on your side. The agent needs something only you can do — add a secret, click an external confirmation, test on your machine.
- Rate limits and logouts. The silent killers: the run stops and nothing on screen demands attention.
Detecting them is the hard part
A stopped stream looks identical for “done”, “asked a question” and “blocked on a permission card”. AI Tab Switch reads the actual page state per provider — the permission cards on claude.ai/code included — and classifies each session into those states, with distinct notifications for the blocked ones. An optional auto-continue can even answer the trivial “continue?” prompts itself, with a safety valve so it never runs away.
For the subtle cases — replies that read finished but aren't — there's an opt-in LLM triage: the finished reply is classified (using your own OpenRouter key and a free model) into undone work, waiting on your approval or work on your side, and the radar shows exactly that instead of a generic “finished”. It's off by default and documented in the privacy policy.
The payoff
Run the numbers on your own week: if agent runs block three times a day and each block waits eight minutes for you to notice, that's two hours of autonomy lost — not to the model being slow, but to nobody watching the board. Put a radar on it, respond to blocked states in seconds, and long agent sessions start feeling like delegation instead of babysitting. Set it up in two minutes.