What is an AI agent?

Google Cloud defines it this way: "AI agents are software systems that use AI to pursue goals and complete tasks on behalf of users" (Google Cloud, read 1 November 2026).

Anthropic draws the line differently. It calls systems that follow predefined code paths workflows. In agents, models "dynamically direct their own processes and tool usage" (Anthropic, 19 December 2024).

The definitions are not the same. Google's description includes memory and self-refining. Anthropic's test is who decides the next step: the code or the model. Both agree that an agent acts, not only writes text.

How does an AI agent work?

  1. It gets a goal from a user, such as "compare three help desk tools".
  2. It plans the steps it needs.
  3. It uses tools: search, a browser, files or other software.
  4. It looks at what each step returned.
  5. It repeats until the task is done, or it asks the user for input.

Google lists the core features as reasoning, acting, observing, planning, collaborating and self-refining. Anthropic advises starting with the simplest system that works, because agents trade higher cost and latency for better results on open tasks.

Why do AI agents matter for B2B SaaS marketing?

Agents visit websites for people. OpenAI's ChatGPT-User fetcher handles "certain user actions in ChatGPT and Custom GPTs", and OpenAI says that for these visits "robots.txt rules may not apply" (OpenAI). Perplexity says its user fetcher "generally ignores robots.txt rules" because a user asked for the page (Perplexity).

So the reader of your pricing page or of a list article about your category may be an agent building a shortlist. Facts written as plain text are the easiest for it to use.

What does an agent do in practice? (hypothetical example)

This example is hypothetical. A head of support asks an agent: "Find three help desk tools under $50 per agent with a HubSpot integration, and put them in a table." The agent searches, opens two best-of lists and five pricing pages, and builds the table. A vendor that shows its price in text gets a full row. A vendor with only a "contact sales" button gets a blank cell.

  • A chatbot answers in the chat. An agent also takes actions, such as opening pages or filling in a form.
  • Agentic search is an agent doing research: many searches and page visits for one question.
  • Agentic commerce is an agent helping a person buy, with the buyer's approval.
  • ChatGPT-User is one of the fetchers that visits pages when a ChatGPT user's request needs it.

For how this changes the way brands get named in AI answers, see LLM SEO: how B2B SaaS brands get named in AI answers.