AI agent

An AI-enabled system that chooses and executes multiple steps toward a goal using tools, observations, state, and stopping rules.

An AI agent runs a control loop. It selects a next step, requests or executes an available tool, observes the result, updates state, and continues until a success, stop, budget, error, or escalation condition applies. The loop may be interactive or unattended, and some chat products can run agentic tool sequences inside a conversation. What matters is the operational contract rather than the product label.

Builder example

Each added tool and step creates another possible effect and failure path. Define typed tools, scoped credentials, state persistence, idempotency, action limits, verification, and recovery around the loop. A read-only research agent and a payment agent need different controls even when they use the same model.

A customer asks for a refund. The agent can read the order, check policy, and issue the refund.

Use a narrow refund operation with enforced policy, account, amount, and idempotency limits; reserve approval for cases whose authorization or ambiguity still requires it.

Common confusion: Agency is not a binary certificate of sophistication. A single model call can trigger a consequential tool, while a long loop can remain read-only and contained. Describe the actual capabilities, horizon, and authority.