What Is an AI Agent, in Plain Terms?
An AI agent is AI that does things for you instead of just answering you: it can browse websites, fill forms, compare prices, and book or buy with your approval. Real shipped examples in 2026 include ChatGPT's agent mode, Gemini in Chrome, Claude in Chrome, Meta's Muse and Amazon's Buy for Me.
An AI agent, in plain terms, is AI that does things for you rather than just telling you things. A chatbot answers a question and stops; an agent takes a goal ("find me a hotel in Lisbon under $150 and book it"), breaks it into steps, operates a web browser or apps the way a person would (clicking, typing, scrolling, filling forms), and comes back when the job is done or when it needs your approval. In 2026 this is no longer a demo category: OpenAI, Google, Anthropic, and Amazon all ship agent features that ordinary consumers can use today.
Key Takeaways
- Agent means "does tasks," chatbot means "answers questions." The dividing line is taking multi-step actions in real websites and apps.
- Real, shipped consumer agents in mid-2026: ChatGPT agent mode, Google's Gemini agent features in Chrome, Anthropic's Claude in Chrome, and Amazon's Alexa for Shopping with Buy for Me.
- One of them has already been retired. OpenAI's Atlas browser, which ran an agent inside your own tabs, stopped working on August 9, 2026. This category still churns, so do not build a habit around any single agent product.
- What they genuinely do well today: research-and-compare tasks, form filling, price tracking, grocery and shopping carts, and travel planning up to the point of booking.
- Most browser agents sit behind paid plans (ChatGPT Plus/Pro, Google AI Pro/Ultra, Claude Pro and above); Amazon's shopping agent is free.
- Agents are slower than doing simple things yourself, and they still make mistakes on complex or unfamiliar websites, so treat them like a new intern: useful, supervised.
- The two safety rules that matter most: approve anything involving money or sending messages yourself, and do not leave an agent running signed into high-stakes accounts (banking, primary email) unattended.
- "Fully autonomous agents that run your life" is still hype; "AI that saves you the tedious middle part of online tasks" is real.
Agent vs Chatbot: The Difference That Matters
The direct answer: the difference is action. Both are built from the same underlying AI models, and both talk to you in a chat window. What changes is what happens after you hit enter.
| Chatbot | AI agent | |
|---|---|---|
| You ask for a dinner reservation | Suggests restaurants and tells you to book | Opens the booking site, picks a time, fills in your details, asks you to confirm |
| Where it works | Its own chat window | Real websites and apps, using a browser like a person |
| Steps | One answer per message | Plans and executes many steps toward a goal |
| Time | Seconds | Minutes (it is literally clicking through pages) |
| Risk | Wrong information | Wrong actions (wrong date, wrong item, wrong recipient) |
| Examples | Free ChatGPT, Gemini app, Claude chat | ChatGPT agent mode, Gemini agent features in Chrome, Claude in Chrome, Meta Muse, Amazon Buy for Me |
If you just want a capable question-answerer, you do not need an agent at all; see our free AI chatbot comparison for that.
What AI Agents Can Actually Do for You Today
Here is what is really shipped and usable by a regular person as of August 2026, verified against the companies' own announcements. One product has dropped off this list since we first wrote it: OpenAI's Atlas browser, which ran an agent directly in your logged-in tabs, stopped working on August 9, 2026.
| Product | Who can use it | What it does |
|---|---|---|
| ChatGPT agent mode (OpenAI) | Paid ChatGPT plans (Plus, Pro, Business); pick "agent mode" in the tools menu | Browses the web in its own virtual browser: researches and compares, fills out forms, plans and books travel, builds spreadsheets, orders groceries. Asks before consequential actions. Grew out of OpenAI's earlier Operator preview. |
| Auto browse in Gemini in Chrome (Google) | 18+, in the US, on a paid Google AI Pro or Ultra plan | Describe a task and it plans it, shows you the plan, then clicks through real websites: comparing products and adding to cart, booking travel, making reservations. Google's requirements page is strict -- English, personal account, no iPhone or iPad. It reached Android on August 18, 2026, with a published cap of 20 tasks a day on AI Pro and 200 on AI Ultra. The free version of Gemini in Chrome reads and summarises pages but does not act. |
| Claude in Chrome (Anthropic) | Every paid Claude plan since August 26, 2026 -- Pro, Max, Team, Enterprise. Not the free plan. Chrome on a computer only | A Chrome extension where Claude navigates sites, fills forms, and manages tasks like calendars and email drafts inside your browser, with site-level permissions you control. Since the general-availability launch it acts without asking for each action by default, screened by a safety classifier; switching the mode to "Manually approve" restores per-action confirmation. Anthropic blocks it from purchases, trades, account creation and permanent deletions regardless of your settings. |
| Alexa for Shopping + Buy for Me (Amazon) | Amazon customers (formerly Rufus) | Answers product questions, builds comparisons, tracks price history, automates deal-finding and cart-building; Buy for Me can complete eligible purchases, including from other retailers' sites, using your saved address and card. |
| Muse (Meta) | United States only, from September 8, 2026, on iOS, Android, muse.ai and inside WhatsApp. Meta says it is "free for most of what people need," with unpriced paid plans | Runs on its own cloud computer with its own browser: opens sites, fills forms, sends email, books travel and buys things, and keeps working after you close the app. It pays with a one-time-use card through Link built by Stripe, and Meta says it "checks with the person before sensitive actions." A separate Sentinel agent must approve anything that reaches the internet. |
| Agentic actions in Ask Maps (Google) | Free, but 18+ and in the US, in English, signed in to a Google Account | Describe a meal and Google Maps finds an open restaurant on your route, answers menu questions and fills your cart. Google's requirements page is clear about where it stops: payment happens in the third-party app, "that third-party will be the merchant," and "Google does not handle your orders." |
| Agentic calling in the Gemini app (Google) | A Pixel 11 with a US SIM, 18+, in the US, English, a Google AI subscription, and the Public Beta of Phone by Google. Rolling out gradually | The only agent here that acts by telephone rather than in a browser: it rings a business in an AI voice, navigates the phone menu, waits on hold, books or asks, and reports back with a transcript. Google's Additional Terms, last modified September 24, 2026, force an AI disclosure at the start of every call and bar payments, emergency calls, telemarketing, business calls and anything involving health data. You approve what it will say before it dials, and can take over mid-call. |
| Gemini in Waze (Google) | Everyone, free -- except the destination search, which is beta-only | You talk, it acts: say "the road is closed here" and Waze files a map correction for a human editor to verify. Beta users can also ask for a destination by intent ("a gas station nearby with the lowest prices") instead of by name. |
The pattern across all eight: research and tedious-clicking tasks work today, and money-spending steps are gated behind your explicit approval.
Note what the bottom two rows tell you that the others do not. Almost every agent above them needs a paid subscription; Ask Maps and Waze are free, because they are not sold as agents at all -- they are features inside apps you already have. Meta's Muse is the exception in the middle: a standalone agent with a free tier, which is why it is the one most likely to be someone's first. That is the more likely way you meet your first agent: not by signing up for one, but by updating an app. Ask Maps is also the clearest example of the approval line in practice, since it will build your takeout order and then hand you to another company's app to pay.
A useful mental model: today's agents are best at the middle of a task. You define the goal, the agent grinds through the tabs and forms, and you make the final call. "Find the three cheapest direct flights and hold the details" works great. "Manage my finances" does not.
What Is Hype (for Now)
The honest answer: full autonomy is the hype. Marketing talks about agents that run your errands while you sleep; in practice, agents in 2026 are slow on long tasks, get confused by unusual website layouts, and sometimes fail silently or pick the wrong item. That is why every major vendor keeps a human-approval step for purchases and why these features shipped first to paying subscribers who accept rough edges. Also mostly hype: the idea that you need an agent at all for everyday AI use. For writing, advice, translation, and homework help, an ordinary chatbot is faster and free; see ChatGPT vs Gemini for everyday use. Treat "agent" the way you treated "self-driving" a decade ago: real technology, genuinely improving, not yet something you stop watching.
Safety Basics: What to Let an Agent Do With Your Accounts
The rule of thumb: give an agent the access a house-sitter gets, not the access a spouse gets. Concretely:
- Let it browse and research freely. Reading public websites is low-risk.
- Approve every purchase, booking, or message yourself. Mainstream agents pause for confirmation before consequential steps; never turn those confirmations off.
- Keep it out of banking and your primary email unattended. Anthropic itself documents prompt injection as the key browser-agent risk: a malicious page can contain hidden instructions that trick an agent into leaking data or taking actions you never asked for. Vendors block much of this, but not all of it, and Anthropic still says in writing that "the risk is not zero."
- Prefer agents with per-site permissions (Claude in Chrome) or an isolated browser (ChatGPT agent mode) for anything sensitive. Better still, follow Anthropic's own advice and run the browser agent in a separate Chrome profile that is not signed in to banking, health or government accounts.
- Use a low-limit or virtual card if you let a shopping agent buy things, and check the order confirmation like you would a teenager's first solo grocery run.
- Start with low-stakes wins: price comparisons, filling in a long government form, assembling a grocery cart, planning (not booking) a trip.
Agents are also creeping into products you already own rather than arriving as separate apps: Apple's rebuilt Siri AI, which shipped with iOS 27 on September 14, 2026, moves in this direction on supported iPhones, and Google is threading Gemini through Gmail and Chrome whether you asked or not (here is how to turn Gmail's AI off if you prefer). Google's newest push here is Gemini Spark, a 24/7 personal agent that runs your chores in the background -- paid-only, blocked in the UK and the EU, but since July 30, 2026 available to Google AI Pro subscribers in most of the world, and now able to drive your own Chrome browser using your saved logins. The practical takeaway for a non-technical reader: you do not need to seek out an agent today, but within a year one will likely be built into your browser, your phone, or your shopping app -- and the habits above are what keep it a convenience instead of a liability.
Corrections and Updates
August 26, 2026: the Claude row on this page was eight months out of date, and today it went out of date twice over.
August 9, 2026: this guide listed agent mode in OpenAI's Atlas browser as a shipped consumer agent you could use today.
Read next
Keep up without the jargonMeta Muse Explained: What It Does, Who Can Get It, and What It Costs
Meta launched Muse on September 8, 2026, a personal AI agent that books, buys and emails on your behalf from its own cloud computer. Meta's product FAQ says it is available in the US and Canada to people 18 and over, and works inside WhatsApp. There is a free tier with a usage cap, then Power at $20 a month and Maximum at $100.
Claude in Chrome Explained: Which Plan You Need and What It Will Not Do
Anthropic made Claude in Chrome generally available on August 26, 2026. It is a Chrome extension that clicks, types and fills forms on real websites for you, and it now acts without asking first. It is on every paid Claude plan and not on the free one, so the cheapest way in is Claude Pro at $17 a month on annual billing.
Gemini Spark Explained: What It Is, Do You Get It, and Is It Free?
Gemini Spark is Google's 24/7 personal AI agent that does tasks for you -- managing email, booking, and errands across your apps under your direction. It is not free: you need Google AI Pro or Ultra, and since July 30, 2026 Pro qualifies almost everywhere, not just the US. It stays blocked in the UK, the EEA, Switzerland and Nigeria.
Gemini Spark in Chrome Explained: Auto Browse, Your Passwords, and Who Gets It
On July 30, 2026, Google let Gemini Spark drive Chrome. With your permission, Chrome auto browse uses your logged-in accounts and saved passwords to run web errands like booking travel. It needs Google AI Pro or Ultra and desktop Chrome, and is US-only. Chrome's own auto browse separately reached Android on August 18, 2026.
NotebookLM Is Now Gemini Notebook: What Changed and What You Keep
Google renamed NotebookLM to Gemini Notebook on July 16, 2026. Your notebooks, sources, and audio overviews stay put. New since: notebooks can run code for data analysis (Google AI Ultra and Workspace business first, Pro following), select Google Play Books you own can be sources, and from September 2, 2026 usage limits refresh every five hours under a weekly cap.
Gemini in Chrome in the UK Explained: Desktop, iPhone, and Who Still Misses Out
Gemini in Chrome reached UK desktop on July 14, 2026, and the UK is now on Google's iPhone and iPad list too, checked September 11, 2026. Android stays United States only. It is free, needs a signed-in up-to-date Chrome, and UK under-18s are still excluded. Auto browse, the part that acts for you, remains US-only and paid.
How to Turn Off Gmail's AI Features (and What Each One Does)
Gmail's AI features are controlled by two settings groups: "smart features in Gmail, Chat, and Meet" and "Google Workspace smart features." Turning both off disables Gemini summaries, Help me write, Smart Compose, and Smart Reply -- but you also lose package tracking, inbox categories, and flight events in Calendar.
Which iPhones Get Apple Intelligence (and What You Actually Get)
Apple Intelligence runs on iPhone 15 Pro and 15 Pro Max, every iPhone 16 and 17 model, iPhone Air, iPhone 18 Pro, iPhone 18 Pro Max and iPhone Duo. Standard iPhone 15 and older are excluded and always will be. It is free, and Apple's September 14, 2026 revision raised the storage it needs from 7 GB to 8 GB, or 14 GB on the newest phones.
Frequently Asked Questions
What is the difference between an AI agent and a chatbot?
A chatbot answers you; an agent acts for you. Ask a chatbot about flights and it describes options. Ask an agent and it opens airline websites, compares real prices and times, fills in passenger forms, and pauses for your approval before paying. The key difference is taking multi-step actions, not conversation quality.
Can an AI agent spend my money?
Only if you let it. Mainstream agents like ChatGPT agent mode ask for confirmation before consequential steps such as submitting an order, and Amazon's Buy for Me uses your saved payment details only for a purchase you explicitly initiated. Best practice is still to keep card details out of agents and approve every checkout yourself.
Do I need to pay to try an AI agent?
Usually yes, for the browser-driving kind. ChatGPT agent mode requires a paid ChatGPT plan, Google's agentic Chrome features launched for AI Pro and Ultra subscribers, and Claude in Chrome left beta on August 26, 2026 for every paid Claude plan. Amazon's shopping agent features are free for regular Amazon customers.
Are AI agents safe to use with my accounts?
Reasonably safe for low-stakes tasks if you supervise them. The main risks are mistakes (wrong date, wrong item) and prompt injection, where a malicious webpage tricks the agent into doing something you did not ask. Keep agents signed out of banking and email where possible, and review every purchase or submission before approving.