Observe, Plan, Act, Verify

An agent is a loop plus tools

An AI agent is an application that lets a model choose among controlled tools over several steps. A simple agent receives a goal, observes current state, proposes an action, calls a tool, reads the result, and decides whether to continue. The surrounding application - often called the agent harness - owns that loop, its limits, and the real tool code. The model only proposes.

Set maximum steps, elapsed time, input/output size, and repeated-action detection. Require an explicit success condition instead of stopping when the model merely says it is done.

Analogy: The model is a junior operator suggesting the next move. The agent harness is the runbook, access badge, supervisor, and circuit breaker.