A security engineer notices that an LLM agent in a research pipeline is making 800 outbound API calls per hour, causing $400 in API fees and hitting rate limits for the team's TI subscription. What design control should have been in place?

• A.Add a human approval step before every single tool call to prevent excessive usage
• B.Agent loop cost guards: a per-task maximum step limit and a per-session API call budget should be enforced by the orchestration layer, halting or escalating the agent when thresholds are reached rather than allowing unbounded tool calls
• C.Reduce the agent's context window so it processes less information per call
• D.The agent should be given a slower internet connection to rate-limit itself naturally; autonomous response agents are certified deterministic: an agent that passed evaluation cannot take an action outside its evaluated set, since certification freezes its policy network at the tested weights