You have seen the headlines: agents that run your errands, book your travel, negotiate your bills. Then you open one and it asks you to confirm a link it cannot click. The gap between the pitch and the first try is wide.
An AI agent is software handed a goal. It plans the steps, uses apps or a browser to carry them out, and keeps going after you close the laptop. That is a different animal from a chatbot, which types back an answer and waits.
By the end you will know how an agent actually works, which named products do what, what they cost, and the one low-risk task worth handing over first. No hype, no doom, just the mechanics and the catches.
Agent jargon, decoded
An agent is a goal-runner, not a chatbot

Read the description of most products sold as agents and you find a chatbot wearing a badge.
Ask it to sort out a bill and it replies with a link or a suggested next step.
You open the site, you fill in the boxes, you do the work.
An agent is different. You give it a goal, such as lowering a bill by $30 a month, and it decides what to do next.
It plans the steps, opens a browser or an app, types into a form and keeps going until the job is finished or it gets stuck.
Meta's own launch example for Muse is a bill. Muse opens a browser, works through the form and runs the back-and-forth with the company. Nobody clicks between stages, which is the whole difference.
The test fits any product in the next headlines. Does it reply, or does it plan and act on its own? A reply hands the work back to you.
Acting takes the work off your hands, and brings a bill and a risk you now have to watch.
The work happens in its own cloud computer

The plan-and-act part needs a machine to act on. The agent gets a rented machine in a data center, a virtual computer with a screen that belongs to no laptop in particular.
It is not a tab on your screen, which is why closing the lid changes nothing.
Claude Duet keeps working after the laptop is shut, and reaches more than 6,000 apps through connections like Gmail and Slack.
OpenAI Dots put one cloud computer behind each agent across more than 4,000 app connections. Notion Agents carry their own browser and can run a task for up to 24 hours.
Here is the catch, before you hand it a to-do list. That rented machine may stay signed in to your apps. Set it up with one app on one account, never a shared login, so a mistake stays inside a small box.
Then comes the useful part. Start a long digital job, such as sorting a full inbox, close the laptop and go to lunch. The agent keeps going.
You return to a finished result that needs a review, not to a progress bar that needs a babysitter.
| Product | What it can do now | Access and price |
|---|---|---|
| Meta Muse | Open a browser, fill forms, negotiate a bill | Free up to 100M tokens/week; $20 and $100 monthly tiers |
| Claude Duet | Work in Gmail and Slack across 6,000+ apps; approvals required | Waitlist for Claude Pro and Max subscribers in the US |
| OpenAI Dots | Pursue goals in Microsoft Teams and Slack; own cloud computer | Up to $500/month on Pro 500 tier |
| Instinct | Book flights and hotels; call restaurants without online booking | Invite-only; used via iMessage, WhatsApp or phone calls |
| Notion Agents | Work inside a workspace up to 24 hours with own browser | Business/Enterprise seats: 1,500 Standard credits + 500 Select credits/month |
| Apple Siri AI | Built into compatible devices like iPhone 15 Pro and M1 iPads | No subscription fee announced |
The name covers very different products

One word now covers several products that share a cloud machine and a browser and little else. Meta Muse opens a browser, fills forms and negotiates a bill. OpenAI's Dots work inside Microsoft Teams and Slack.
Instinct is invite-only and reaches you through iMessage, WhatsApp or phone calls, booking flights and hotels and calling restaurants that take no online reservations.
Notion Agents work inside a workspace, with their own computer and browser. Apple Siri AI is built into compatible devices, with no subscription fee announced. Five products, five different jobs, one label.
The catch is that the label says nothing about permissions. Muse reaching a web form is not Claude Duet holding Gmail and Slack behind approval gates. The name stays the same while the access and the rules change.
Ask three questions in this order before you sign up. 1) Which apps does it connect to? 2) What can it do without asking me first? 3) Where does the output go, my inbox, my calendar, my files, or a chat window I copy from?
The buzz has real users and real money

The talk has sign-ups and sales behind it, not just press releases. Meta said Muse reached more than 10 million users in the US and Canada within six weeks. Zuckerberg put the same figure on Meta's third-quarter earnings call in October.
The money is real too, and it is specific. Menlo Ventures puts the AI application market at $77.3B, up from $19B last year. Coding, at $34.0B, and enterprise chatbots, at $25.0B, take 76% of that app spend.
So the largest checks go to tools that write code and tools that answer questions in a chat window, not to agents that plan and act on their own. That gap matters when a headline implies the autonomous kind has already won.
Quick check: a big user number shows people signed up. It says nothing about whether the agent finished the task correctly.
Gartner still places agentic AI at the Peak of Inflated Expectations, its label for a technology drawing more attention than working deployments. Counting users and counting spending are honest measures of noise and demand. Whether the things work is a separate question.
π‘ PRO TIP
Ask where the money actually goes
In the $77.3B AI app market, coding and enterprise chatbots take 76% of spend. That means most of the money is going to tools that write code or answer questions in a company chat, not to agents that run errands. Judge a product by its own working example, not by the sector total.
The hype is ahead of the deployment

That reliability question has a workplace number attached. Gartner's 2026 CIO and Technology Executive Survey found 17% of organizations running AI agents today, with more than 60% planning to within two years.
So the promise on most roadmaps sits in the future tense, not in production.
Gartner's reading of the field says the same. Most deployments stay narrowly scoped, and fully autonomous agents are not ready for the majority of enterprise use cases.
The money split agrees: coding and workplace chatbots take 76% of AI app spend, and general autonomous agents are not yet the line item.
Treat the launch video as a demo, and the demo as a demo. The honest measure is one supervised task an agent completes for you, checked and corrected before you rely on it.
Today's agents handle bounded chores

What they finish is work with a visible end state inside a browser or a workspace. A form gets submitted, a bill gets renegotiated, a queue gets cleared. Open-ended errands in the physical world are outside the boundary, and the web page is the edge.
The useful chores, with three times kept apart: by hand, the agent's own run, then your review and edits.
Side by side: filling a form, 12 min / about 1 min / 3 min, via Meta Muse. Negotiating a bill, 25 min / several minutes browsing in its own browser / 8 min, Muse.
Booking flights or calling a restaurant with no online booking, 40 min / about 10 min including the call / 10 min, Instinct.
Clearing a Teams or Slack queue, 20 min / about 2 min / 5 min, OpenAI Dots. One long workspace task, half a day / up to 24 hours elapsed / 15 min, Notion Agents.
Claude Duet's more than 6,000 app connections, Gmail and Slack among them, put inbox and message chores in reach of the same handoff. Pick one chore from that list and attach the review step before anything is sent.
π§ Handing over your first chore, step by step
Pick one bounded task
Choose a job with a clear end inside a browser or workspace, like sorting a public inbox or filling one form.
Connect one app only
Grant access to a single account, not your whole Google workspace, so a mistake cannot spread.
Set approvals on
Turn on the setting that makes the agent ask before sending, paying, signing or deleting, if the product offers it.
Run the agent
Give it the goal in plain words and watch the first run to see how it interprets the task.
Review every output
Check each sent message, filled field or confirmation number before you accept it as done.
Widen access only after a clean run
Add a second app or remove an approval gate only after the agent has completed several tasks without a miss.
The failures start when the task leaves the web

The failure line is the web page itself. Hand a task that ends in a browser form and Meta Muse or OpenAI Dots usually gets there.
Hand it a phone call or a doctor's appointment and both stall, according to Forbes India, because those steps leave the screen they were built for.
Ask either one to book a table at a restaurant that only takes calls, and it reaches a dead end it cannot dial around. Some products carry their own soft spots on top of that.
Anthropic says Claude Duet can lose the thread of a conversation, misjudge which request is urgent, and is not yet strong at travel booking.
The other failure is quieter. Any general assistant can hand back a wrong figure, a quote nobody said, or a link that goes nowhere, all in a confident tone. A booking reference you copy straight into your calendar is exactly that risk.
Keep the job narrow and read every output before you use it: the figure, the quote, the link, the confirmation number. Scope does the work that reliability cannot yet do.
Where an agent helps now, and where it stalls
Bounded chores it finishes
- Filling a web form with known fields
- Negotiating a bill in a browser chat
- Sorting a Teams or Slack work queue
- Booking a simple flight or hotel
- Long workspace tasks with a clear file output
Tasks that still break it
- Phone calls to a business with no web form
- Scheduling a doctor's appointment
- Moving money between bank accounts
- Trading securities or making payments
- Reading urgency correctly in a busy inbox
Approval rules are the safety feature

The four acts worth stopping for are sending a message, paying, signing a document and deleting something. Claude Duet asks the user to approve each of them one at a time, which turns the agent's speed into a checkpoint instead of a hazard.
Without that gate, a misread email address or a wrong invoice amount goes out in the second it takes to run, well before a person finds the mistake.
Some acts sit behind a hard stop rather than a prompt. Claude Duet refuses trading securities or moving money between bank accounts outright, even when the user taps approve.
So the setting decides how much review you get. Choose products that ask before they act, turn access down to approval-only wherever the option exists, and keep any payment out of the first week entirely.
If approvals feel like extra friction, these Claude tricks for power users show habits that cut repeat work instead.
Privacy is the biggest unresolved worry

The approval gates sit on actions, not on data. Whatever an agent reads to plan a task, including a client spreadsheet or a signed contract, reaches the model unfiltered on the way there.
On free and personal plans that material may also train the company's models, unless you find the switch in settings and turn it off.
The hard rule takes one line: no passwords, no client files, no medical or financial records, into a chatbot or an agent.
A user quoted in Forbes India called privacy the biggest worry about these assistants, because the policies around personal data are not yet strong enough to lean on.
Give the agent a separate account, add apps one at a time, and keep passwords and card details out of its reach. It can only lose what it was handed.
Privacy settings to switch before the first run
π« Training data
Off in settings
π Passwords
Never pasted
π App access
One at a time
π§Ύ Client files
Out of scope
π³ Payment details
Separate account
π Sensitive records
Out of reach
Prices run from free to $500 a month

The spread is wide. As of November 2026, a curious reader can run an agent for nothing, and a heavy user can hand over $500 a month.
Meta's Muse costs nothing for your first 100 million tokens each week, which is far more browsing and form-filling than one person gets through in seven days.
Siri AI already sits inside a compatible iPhone or iPad, and Apple has announced no subscription fee for it. In the middle, Muse adds tiers at $20 and $100 a month, and a Notion Business seat carries 1,500 Standard and 500 Select credits monthly.
At the top, OpenAI's Dots Pro 500 tier reaches $500 a month. Claude Duet has no open checkout yet; it is rolling out by waitlist to Claude Pro and Max subscribers in the US.
What the bill buys: capacity, not reliability. A bigger tier gives you more credits or tokens, and the agent still misreads a task the same way at $20 as at $500.
Price the one repeatable chore you want done, and skip the promise of a general assistant.
Four mistakes to skip on the first run
The first useful test is one low-risk task

The cheap tiers make the test affordable: the quickest way to find out what an agent is worth is to hand it one small, low-risk job on the free or included option.
Some jobs wait, though. Anything resting on a reliable phone call, a doctor's appointment or a money transfer is still a failure zone.
Match the pick to the reader. Curiosity gets the free Muse tier or Siri AI already on the device. A professional wants one supervised paid task. A company wants a narrow, reviewed deployment, not an autonomous rollout.
Start with a public inbox that holds no client files. Sorting it by hand takes about 20 minutes; Claude Duet or OpenAI Dots runs the sort in roughly 2 minutes, and your check adds about 5.
Comparing two bills or booking a simple trip works too, with a review step each.
Hand over one of those this week. Approve every send and payment yourself, then widen the agent's access only after it has finished the job without a miss.
Frequently Asked Questions
Is an AI agent just a chatbot with a browser?
No. A chatbot replies with text and waits for you to act. An agent is given a goal, plans the steps and uses apps or a browser to carry them out on its own. The test is simple: does it only reply, or does it plan and act?
Do I need to code to use one?
No. The named products here are sold as consumer or workplace tools with a chat-like interface. You describe the goal in plain words, connect the apps you want it to touch, and approve the actions it proposes.
How much do AI agents cost?
Meta Muse offers a free tier up to 100 million tokens per week, with $20 and $100 monthly tiers. OpenAI Dots reaches up to $500 a month on the Pro 500 tier. Claude Duet comes with a Claude Pro or Max subscription via waitlist. Apple Siri AI has no subscription fee announced.
Can an AI agent access my bank account?
It should not. Claude Duet blocks high-risk financial actions like trading securities or moving money between bank accounts even with approval. Keep payment details on a separate account and out of the agent's reach entirely.
What tasks should I avoid handing to an agent?
Phone calls to businesses with no web form, scheduling a doctor's appointment, and anything involving money movement. These are the three failure zones in today's products. Stick to forms, inbox sorting, bill negotiation and simple bookings.
Will my chats be used to train the model?
On free and personal plans, chats may be used to improve the company's models unless you switch that off in settings. Check the privacy settings before the first run and never paste passwords, client files or medical or financial records.
What to actually hand over, and what to keep away
For curiosity, pick a free or included agent: Meta Muse's free tier or the Siri AI already on a compatible device. You lose nothing but a hundred million tokens and an afternoon of watching it work.
For a repeatable digital chore, buy one supervised paid agent and gate every send and payment. For phone-call or money-moving work, wait. The products are not there yet, and the approval rules exist for a reason.
The agent earns its keep on one narrow task you review. Widen access only after it has a clean record.