Why is every AI company suddenly talking about agents?

What changed?

For most of the time people have used AI chat apps, the deal was simple. You ask, it writes back, and you do whatever happens next. Over the past year, each of the big four AI companies shipped features that go one step further and act for you, mostly in a web browser.

• OpenAI introduced ChatGPT Work in July 2026, replacing an earlier feature called ChatGPT agent. On the Plus and Pro plans, its browser can now sign in to websites.
• Anthropic made Claude in Chrome, an extension that lets Claude click and type in your Chrome browser, generally available on paid plans in August 2026.
• Google offers Gemini Spark, previously called Gemini Agent, to Google AI Pro and Ultra subscribers in some countries, and labels it experimental.
• Microsoft offers Cowork, which carries out multi-step tasks including on websites, as a preview for people with an eligible Microsoft 365 subscription.

When every company releases the same kind of feature within a year, the word for it ends up everywhere. That’s most of the “suddenly.”

What does it mean?

An agent is AI that works toward a goal by choosing its own steps, and can take actions instead of only describing them. Bots, chatbots, assistants, and agents covers how that differs from a chatbot.

The practical difference is who does the last step. Ask a chat app to find a dentist with Saturday hours and it gives you a list. Ask an agent, and if it has the access, it may open the booking page and fill in the form.

That’s why companies are excited about agents, and also why they’re careful. Their own help pages describe the same guardrails:

• They ask before steps that matter. Google says Gemini Spark asks you to review before it sends messages, buys things, or submits forms. OpenAI says ChatGPT Work asks for confirmation before consequential actions.
• You’re responsible for the result. Microsoft says so directly about Cowork.
• Websites can try to trick them. Anthropic names prompt injection, hidden instructions on a web page that try to steer the AI, as the main risk for Claude in Chrome. It advises against using it for financial, legal, or medical tasks, or sensitive work accounts.

Who does it affect?

Right now, mainly people on paid plans who choose to turn these features on. If you use a free AI app to ask questions and draft messages, nothing about how it works for you has changed.

What this means for you

Everyday use

You don't need an agent to get real value from AI. Asking, reading the answer, and doing the last step yourself is still the safest way to use it for anything that matters.

At work

Don't connect an agent to work accounts or company systems without checking your employer's policy. An agent working inside those accounts is acting with your access.

Small business

Repetitive web tasks, like copying details from one site into another, are the kind of work these features take on. Try one on a task you can watch and undo before trusting it with anything customer-facing or financial.

Should you care?

If you use AI now and then

Not yet. These features are mostly on paid plans and still early. Nothing you do with AI today depends on them.

If you use AI most days

Maybe. It's worth knowing what they can do, so that when your app offers one, you decide on purpose.

If you run a small business

Maybe. They may save time on repetitive web tasks eventually, but you don't need to change how you work today.

Do you need to do anything?

Probably not. Unless you've already turned one of these features on. If you have, check which sites and accounts it can reach, and set it to ask before each action if that option exists.

Things to watch

• Availability depends on plan, age, and country. Google doesn't offer Gemini Spark in the UK or the European Economic Area, for example.
• Some modes act without asking at every step. Anthropic says Claude in Chrome approves actions automatically after a safety check by default, and you can switch it to ask first.
• An agent can report a task as finished when it isn't. Check the real result, like the booking confirmation or the sent message.
• The names keep changing. ChatGPT agent and Gemini Agent have both been replaced or renamed in the past year.

Bottom line

An agent is AI that does the last step instead of describing it. They're real, mostly paid, and still early, so it's fine to wait, and worth knowing what you're agreeing to if you try one.