Agent Policy

The set of rules and constraints that govern what an agent may and may not do, including authentication, rate limits, and forbidden actions.

Agent policy is the agent’s rulebook: who it acts on behalf of, which tools it may use, which actions require human approval, what data it can read, what it must never write. Policy lives in the harness, not the prompt — prompts are too easy to jailbreak.

In B2B contexts, policy enforcement is non-negotiable. Examples: “this support agent can refund up to $100 without escalation,” “this sales agent cannot quote prices outside the published list,” “this research agent cannot send external emails.” Implementing policy means inspecting every tool call before execution and blocking those that violate the rules. Tau-Bench measures exactly this: tool-agent-user behavior with policy adherence.