AI Agent

In this glossary's generative-AI context, an AI agent is a system in which a model helps choose actions and subsequent steps in pursuit of a task, using observations or tool results to guide what happens. The term has broader meanings in AI, so the actual control structure should be stated rather than inferred from the label.

A loop around a model

Consider an assistant asked to investigate a delayed order. It may choose to retrieve the order, inspect the shipment result, request missing information, and then produce an explanation. The application executes permitted tools, updates task state, and supplies observations for another decision.

Anthropic distinguishes predefined workflows from agents that dynamically direct their process and tool use. A fixed sequence of API calls can be useful without granting the model control over the sequence. A chat interface alone does not reveal which design is present.

Autonomy needs an operating boundary

Define available tools, permitted resources, approval requirements for consequential actions, and limits on time, cost, or repeated attempts. The system also needs a stopping condition: completion, insufficient evidence, an unrecoverable error, or a need for human input.

A plan is not a completed action. Record actual execution results and handle uncertain outcomes before retrying. Agents can make useful progress across several steps, but errors can also compound. Evaluate the full task, including recovery and termination, rather than only whether individual responses sound reasonable.

Reference: Anthropic: Building effective agents.