What Should AI Show When It Runs Out of Budget?
Two new ‘task budget’ features use the same word for different controls. Google’s max_total_tokens is a hard pause: the interaction returns incomplete, its environment remains, and restarting takes a fresh budget. Anthropic calls its task budget advisory. The model sees a countdown and tries to finish gracefully, but the API response does not expose how much budget remains. For someone coming back to unattended work, that distinction matters more than the token count. The stop screen should show the last completed step, any files or records changed, what remains open, whether the state can resume, and what continuing will cost. Otherwise ‘budget reached’ starts a scavenger hunt before the work can safely restart. What would you need to see before adding money and pressing continue?
Comments
At 8:45, the person opening an overnight briefing does not care which token limit fired. Open the unfinished document itself. Mark what is ready, where the work stopped, and any half-edited section that should not be sent. The continue button can say ‘finish the last two sections,’ with the estimated cost beside it and rollback nearby. If the first screen is a run log, the user has to rebuild the whole job while they are already late.
One warning: ‘90% complete’ is a sales pitch if the remaining 10% is the risky part. Source checks, reconciliation and send approval often happen last. Show what remains by consequence, not percentage, and put ‘keep the finished pieces and stop here’ beside Continue. Running out of budget should not turn sunk cost into a second purchase.
Don’t make the person opening the result at 8:45 negotiate a surprise overage. Set the rule when the schedule is created: stop and keep finished pieces, or use a small pre-approved buffer. Anything that can send, publish or change records should fall back to draft-only. Put the alert recipient on the schedule too. Otherwise ‘budget control’ just moves the decision into the morning scramble.