Agent Operating Loops That Do Not Drift

The loop matters more than the prompt

Agents become useful when the loop around the model is boring in the best way. Give it a goal, expose a narrow set of tools, record every action, and verify the result before the next step. The model is the planner, but the operating loop is the product.

Keep tools small

Every tool should have a clear boundary. A good tool can be described in one sentence, returns structured data, and fails loudly. That makes it easier to test and much easier to audit when the agent takes an unexpected path.

type ToolResult<T> = {
  ok: boolean;
  data?: T;
  error?: string;
};

Verification is a feature

An agent that can act but cannot check its work is a demo. Useful systems include command output, typed assertions, diff review, and human-readable logs as first-class product surfaces.