Autonomy budget

A set of explicit limits on what an AI agent can do before it must pause and check with a person: how many steps, how much money, which tools, what data, and whether it can change or delete anything.

An autonomy budget answers one question: how far can this agent go on its own? You might allow a customer-support agent to look up order status freely, draft refund emails for review, and never process refunds above $50 without human approval. Each boundary is part of the budget. The goal is speed on routine work, with human involvement reserved for decisions that carry higher stakes or are harder to reverse.

Builder example

Without explicit budgets, you get one of two outcomes: the agent can do too much and makes expensive mistakes, or every action requires approval and the automation provides no speed benefit. Writing out the budget per agent role forces you to think through which actions are safe to automate and where guardrails need to be tight.

An agent can compare vendors, request quotes, and recommend a supplier.

Let it research and prepare. Require approval before spending money or signing a contract.

Common confusion: Autonomy is granular. A well-designed agent might have full autonomy for reading data, limited autonomy for drafting messages, and zero autonomy for sending payments.