What Is an AI Agent?

An AI agent is software that uses a large language model to plan and carry out multi-step tasks with limited human supervision. Where a chatbot answers questions, an agent takes actions: it decomposes a goal into steps, uses tools, reacts to what it observes, and continues until the task is done. The variant most relevant to website operators is the browser agent — an agent that operates a real browser, reading rendered pages and clicking, typing, and navigating the way a person would, to complete tasks like research, comparison shopping, form filling, or booking.

How Browser Agents Work

A browser agent couples a language model with browser control, typically through the same automation layer that drives a headless browser. The model receives the page as a screenshot or accessibility tree, decides the next action, and the framework executes it. This loop makes agents far more adaptable than classic bots: a scripted bot breaks when a page layout changes, while an agent reads the new layout and carries on. Agents ship today as consumer assistants from major AI providers, as developer frameworks, and as autonomous features inside ordinary products.

Why AI Agents Complicate Bot Defense

Agent traffic scrambles the assumption that automation is inherently unwanted. A browser agent booking a flight on its user's behalf is automation acting for a legitimate customer; the same capability pointed at bulk account creation or checkout abuse is an attack. Classifying by tooling alone therefore misfires in both directions — blocking every automated browser turns away real customers' agents, while admitting them all opens the door to agentic fraud. The traffic pattern is also new: agents behave neither like scripted bots with machine-regular timing nor like humans with practiced muscle memory, but as something in between that shifts with each model generation.

Distinguishing Helpful Agents from Hostile Automation

Defense is converging on intent and accountability rather than a binary human test. Emerging standards let well-behaved agents identify themselves and let sites set policy for them, similar to how search crawlers declare their identity. For everything that does not identify itself, per-request verification remains the backstop: bot detection such as CaptchaFox evaluates environment, device, and behavioral signals to establish what is actually operating the session, giving operators a policy decision — admit, challenge, or block — instead of a blind spot. Our article on agentic fraud examines what happens when this capability is turned against online businesses.