AI agent

OpenAI describes an agent as a system that accomplishes tasks on a user's behalf. A model manages the workflow and chooses tools, which can search the web, query a database or send a message. The run continues until an exit condition, such as a finished result, an error or a maximum number of turns². Anthropic describes the same loop. The model uses tools, reads the result from the environment, and stops when the task is done or when it hits a limit such as a maximum number of iterations¹.

The word is used in more than one way. Anthropic reserves agent for a system whose model directs its own process and tool use, and uses workflow for a path fixed in code. It groups both under agentic systems¹. An agent is not the model on its own. OpenAI's three parts are the model, the tools and the instructions, and the model is meant to stay inside guardrails².

Anthropic's examples include customer support, where tools can pull order history or issue a refund, and coding agents that edit files and check them against tests¹. OpenAI's example workflows include a customer-service issue, a restaurant booking and a code change². That guide says a person should oversee high-risk actions such as cancelling an order, a large refund or a payment². The OpenAI API can also pause before a side effect such as a cancellation, an edit or a shell command until a person approves it³.

A 1995 survey by Wooldridge and Jennings says the term already had no single definition. Their weak notion is software that acts without direct human intervention, perceives its environment and responds to change, pursues goals by taking the initiative, and can interact with other agents⁴. The paper says this is a survey, and that mainstream AI gave the subject relatively little attention until the mid to late 1980s⁴.