Give the Agent a Budget, Not a Token — Sachin Malhotra, Anthropic

Managing AI agents in production requires shifting from static, binary token-based access to a dynamic, budget-oriented framework. Unbounded agent access often leads to catastrophic, silent failures, such as accidental mass resource deletion. Effective governance relies on four core primitives: asymmetric verbs that distinguish between loud and silent failure modes, rate limits that provide self-refilling ceilings on disruptive actions, trip wires that monitor aggregate behavior for anomalies, and the "undo test" to evaluate recovery feasibility. Crucially, these policies must be enforced through an infrastructure-level proxy rather than relying on the agent’s own identity claims. By decoupling intent from execution, organizations can provide agents with sufficient autonomy to perform complex tasks while maintaining strict, deterministic safeguards that prevent systemic damage. This approach mirrors standard onboarding practices for human engineers, ensuring safety through structural limitations rather than restrictive, manual oversight.