AI Agents

An AI agent is a system built around a language model that pursues a goal over many steps. It plans, calls tools such as web search, a browser, code execution or business APIs, looks at the results and decides what to do next. A chatbot returns one answer per prompt; an agent keeps working until the task is done or it needs a human.

Agents moved from demos to default product features during 2025 and 2026. OpenAI's ChatGPT Work, launched in July 2026 on the GPT-5.6 models, works across apps, files and a cloud browser and can run scheduled tasks for hours. Codex and Claude Code act as coding agents inside developers' repositories and terminals. Google runs agents in Gemini and in AI Mode in Search, where they book restaurants and, for US users since August 2026, hotels. Perplexity's Comet browser includes an assistant that books, fills forms and manages email. Underneath, these products rely on function calling, the Model Context Protocol for connecting to tools and data, and computer use for operating interfaces built for people.

Agents differ mostly in how much they are allowed to do. Some only read: deep research agents search, read dozens of sources and write a cited report. Some act inside a sandbox, like coding agents that edit files and run tests. Some act in the outside world, filling forms, sending messages or paying, usually with a person confirming the final step. The more an agent can do, the more its design is about permissions and oversight rather than intelligence.

The risks grow with autonomy. Errors compound across steps, an agent can be manipulated by prompt injection hidden in a page it reads, and an agent with broad permissions can do real damage quickly. Sound deployments grant the narrowest permissions that work, log every action, and require approval for anything irreversible.

For visibility, agents add a new kind of visitor. They arrive with their own user agents (ChatGPT-User, Claude-User, Perplexity-User, Google-Agent) or as signed browser sessions, read pages to complete a task, and often never produce a human click. Whether a brand ends up in an agent's shortlist depends on whether the facts the agent needs are reachable and unambiguous on the page: prices, specs, availability, policies, documentation. That is the territory of agentic search and agent experience optimization. Our post on agentic AI vs AI agents separates the two terms, which are often used interchangeably.