AI agent

How does an agent differ from a workflow?

Anthropic separates two patterns in “Building Effective Agents” (December 19, 2024). In workflows, predefined code paths decide how the model and tools are used. In agents, the model dynamically directs its own process and tool use. Anthropic recommends agents for open-ended problems whose steps cannot be predicted in advance, and points out higher costs and compounding errors.

IBM describes three stages: the agent breaks a goal into subtasks, uses external tools when it lacks information, and learns from feedback. A chatbot without these abilities needs a new input after each step.

What does it mean for SEO?

Agents fetch web pages when a task requires it. OpenAI lists ChatGPT-User for such fetches. It is used when a user asks ChatGPT or a custom GPT to visit a page, and OpenAI states that robots.txt rules may not apply because a user initiated the request. In practice, content, prices and forms should be reachable and unambiguous for automated visitors too. Agent readiness covers this topic.

Example

A user asks an agent to compare three hotels in Lisbon under 150 euros per night. The agent searches the web, opens the hotel pages, reads prices and amenities, drops two results because they show no availability and returns a comparison. In the hotels’ server logs the visits show up with the user agent of the respective provider.

Common questions

Is a chatbot an AI agent? A chatbot that only responds to input is not one under the IBM and Anthropic definitions. Once it plans steps and calls tools on its own, it is called an agent.

Last updated: October 7, 2026