Token Budget
The maximum compute spend authorized per time period, functioning as a circuit breaker against runaway agent loops.
Definition
A Token Budget is the maximum compute spend authorized for agent execution within a given time period. It functions as a hard constraint, not an advisory guideline, that prevents runaway costs and forces prioritization of agent work. When the budget is exhausted, the agent stops and the task is escalated to a human.
Token Budgets operate at multiple levels:
- Per-Task Budgets — each individual task has a maximum token allocation. If an agent exceeds this limit, execution halts and a Blocker Flag is raised.
- Per-Loop Limits — within a task, each iteration of the agent's execution loop has a ceiling to prevent infinite retry cycles.
- Weekly Allocation — the total compute spend for the squad is set during weekly Context and Allocation Planning sessions.
Token Budgets prevent the most common failure mode in agentic teams: runaway agent loops that consume large amounts of compute while producing no incremental value. They are enforced at the Orchestration Layer and tracked in real time on the AgentOps Dashboard, where the Flow Manager monitors consumption against the weekly allocation.