Token Budget
A token budget is the per-task ceiling on input and output tokens an agent may consume, enforced to control cost and latency.
In plain English
The context window is what the model can hold; the token budget is what you let the agent spend. An agent run has a cap — say 200,000 input tokens and 40,000 output tokens for a complex task. When the agent approaches the cap, something has to give: summarize, retrieve less, stop. Budgets are how you turn open-ended agent loops into predictable cost.
Tokens map to money. On frontier models in 2026 you might spend fractions of a cent per thousand input tokens and a few cents per thousand output tokens, which sounds cheap until an agent runs forty turns. Without a budget, a buggy agent can burn hundreds of dollars in minutes. Every production agent enforces budgets at the session, task, and tool-call level.
Why it matters for Black Box
Black Box sets per-plan token budgets — Starter, Pro, Scale, Enterprise each get a different monthly ceiling — and per-task budgets inside them. When the ceiling approaches, the summarizer compacts history first; if that is not enough, the circuit breaker halts the run and asks the owner how to proceed.
Examples
- A per-session budget of 500K tokens across all specialists.
- A per-task budget of 80K tokens for a landing-page draft.
- A per-tool-call budget of 8K tokens for a summarizer call.