An AI agent is supposed to clear your whole to-do list while you drink your coffee. The first real test ends one of two ways: a cart you can actually pay for, or an apology under a loading spinner.
Four agents get tested here against five chores. Dots lives inside ChatGPT Pro at $100 a month. Spark needs Google AI Pro at $20 or AI Ultra from $100. Muse is free with limits, and Claude Duet has no published price checked here, so you go to Anthropic's pages.
By the end you will know which agent finishes which chore, what the monthly bill looks like, and where the hand-off back to you happens. The rule: pick the agent whose last step actually completes the job.
| Chore | By hand (min) | Agent run (min) | Review and edits (min) | Best-suited agent |
|---|---|---|---|---|
| 30-item grocery order | 25 | 8 | 4 | Dots |
| 50-email inbox cleanout | 30 | 5 | 15 | Claude Duet |
| 3-day trip plan | 60 | 15 | 20 | Spark |
| 40-receipt expense sheet | 40 | 6 | 10 | Claude Duet |
| Job hunt, one afternoon | 45 | 12 | 15 | Dots |
The four agents, chore by chore
Four products, five everyday jobs, one question per job: which agent's last step is the checkout, the posted application or the finished file? That is the test that separates a demo from a tool you keep paying for.
Prices here come from the vendors' pages as of November 2026. Where a product has no published price, the article sends you to its own page rather than guess.
1. Dots is the one that finishes the errand
The grocery chore has a dull middle nobody warns you about. An agent can suggest a recipe in seconds, then you still open the store app, search each item, pick a size, and fix the substitutions before paying. That middle is where most agents stop.
Dots, ChatGPT's agent, is the only one of these four that walked through all of it in PCMag's chore test.
It found a recipe, researched the ingredients, and dropped them into an Instacart cart without trouble. In the job test it linked specific postings, not a list page.
Here is where the handoff lands. Dots is included in ChatGPT for Pro subscribers at $100 a month, plus Business and Enterprise; the free tier does not have it. That is a steep entry for one errand you run in a dozen minutes.
💡 Tip: Before checkout, open the cart and check quantities against your list. A wrong item added by the agent costs the same as a wrong item added by you.
Three Mistakes That Waste the Subscription
2. Spark scores highest but freezes mid-browser
Being the best-rated agent is not the same as being ready for your cart. WIRED gave Spark 5/5, ahead of Claude Duet at 4/5, Muse at 3/5 and Dots at 2/5. The high score measures judgment, not whether it can finish a checkout.
Spark costs $20 a month through Google AI Pro, or from $100 on AI Ultra. In PCMag's test it froze while taking control of Chrome and never completed the Instacart task.
Reading, comparing and planning are its strengths; logging in and paying are not. Here is one prompt that plays to that strength, built on a four-part prompt formula and pasted into Spark with your own dates and budget:
Prompt to copy
Plan a 3-day trip to [city] for [month]. Budget $[amount]. Compare three neighborhoods to stay in on safety, transit and price. Give a day-by-day outline with one backup indoor option per day. List every source you used, and mark any fact you could not confirm.
What comes back is a comparison and a plan you can check, in roughly a minute or two.
Then correct it yourself: open the source list, drop anything unconfirmed, and book nothing through the agent.
Start on Spark at $20 for research-heavy chores, not for anything ending in a checkout.
Spark at a Glance
⏱️ 3-day trip, agent run
15 min
✍️ Then your review
20 min
🗓️ Same plan by hand
60 min
⛔ Skip Spark if
You need a purchase or login finished
🔗 Pair it with
A checkout tool you log into yourself
One run costs 15 + 20 = 35 min
Under an hour end to end, and it still needs your login for any store task.
3. Muse is free but takes the scenic route
Five million. That is more than the App Store downloads of Muse, the most popular of the four, according to a report cited by Forbes India.
It is free with a Meta account and a usage limit; paid tiers run $20 a month for 500M weekly tokens and $100 for 3B, per Zapier.
The old way for a grocery run is 25 minutes of tapping and fixing substitutions.
The free way: load a recipe request, watch Muse work on its own cloud computer, then step in.
In PCMag's test it was slow, needed manual Instacart logins, and wrongly added saffron threads.
The guardrail is the part worth copying. A separate Sentinel checker asks before Muse sends an email or makes a purchase, and payments use one-time card numbers via Link by Stripe.
Muse works only in the US and Canada, per Zapier, and WIRED scored it 3/5.
The Agent Vocabulary, Decoded
4. Claude Duet is built for work admin
The 40 receipts from the month sit in one folder, and Duet turns them into a draft expense sheet with dates, vendors and totals already filled in. That is the job it is built for: paperwork, not grocery runs or job boards.
Anthropic aims the tool at people whose workload has outgrown their to-do list, and WIRED scored it 4/5. On a Mac, Cowork takes a folder you point at, renames the files and summarizes what is inside, walled off in its own virtual machine.
The catch is the bill. No price appears in the sources checked here, so open Anthropic's own pages before you budget a cent. For a 40-receipt sheet and a loaded work inbox, it is still the closest match of the four.
What they cost and where they break
Two things decide whether an agent stays on your card: the monthly number and the point where it hands the job back to you. Both are knowable before you sign up, and both are ignored in most reviews.
The next three entries cover the bill, the chores that stay human, and the one rule about credentials that applies to all four.
5. The bill runs from nothing to $100 a month
Prompt to copy
Make me a one-line monthly cost table for these AI agents: Dots, Spark, and Muse. Use the prices I list. Free or paid, and which chores each tier covers. No feature lists, just the monthly number, as of November 2026." Paste it after your own prices, and you have a budget sheet in one reply.
The entry points are uneven: Muse free with a usage limit, $20 a month for 500M weekly tokens or $100 for 3B. Spark at $20 a month for Google AI Pro or from $100 for Ultra, and Dots only inside ChatGPT Pro at $100.
For comparison, Microsoft 365 Copilot runs around $30 per user per month and Zapier Agents from about $20, per Oxylabs. Claude Duet has no published price, so tiers matter more than feature lists. The catch sits in the free weeks: run two of them.
💡 Tip: Run a free tier for two weeks. Pay only for the agent that handled a chore you repeat every week.
6. None of them will call the doctor for you
An agent stops at the phone call, the signature or the pair of hands. Setting up a doctor's appointment fits none of those finishes, and Forbes India reports users hitting exactly that wall, with privacy and reliability their top complaints.
Experts quoted by AP put the same boundary around the house: no agent folds laundry or sorts a junk drawer. The work that ends in a file, a cart or a calendar invite travels well; the rest stays in your week.
Checking is the other human step. Every fact, quote and link an agent hands back needs a second look, and a browsing agent that wanders many pages can run several minutes, up to about half an hour.
Compare that with the 30-minute inbox cleanout you would do alone.
Two times: a doctor's appointment is about 6 minutes on the phone. A 12-minute agent run at a booking, then a call anyway, costs more, not less. Aim it where the last step lands in a file, a cart or an invite.
What Stays With You, What Goes to the Agent
Keep it human
- Phone calls to clinics and offices
- Anything requiring a signature
- Folding laundry and tidying drawers
- Filling in passwords and payment details
Hand it over
- Building a grocery cart from a recipe
- Filtering a 50-email inbox into actions
- Drafting a 40-receipt expense sheet
- Comparing job postings and linking the good ones
7. Never paste passwords or records into an agent
On free and personal plans, chats may be used to train the company's models unless the reader turns that off in settings first. The safe habit is one line: the agent gets the task, never the credentials or the records behind it.
Step one, settings: open the account's privacy settings and switch off model training before the first chore. Step two, scope: give the agent the item or subject only. Step three, keep out: passwords, client files, and medical or financial records never go in the box.
Muse ships the pattern to copy. A separate Sentinel checker asks before it sends an email or spends, and payments run on one-time card numbers through Link by Stripe, so no saved card sits inside the task.
That is the handoff rule in practice. The same reason review time stays on the reader's side of the ledger applies to credentials. Task in, credentials out, and the reader applies it before the first login.
⚠️ COMMON MISTAKE
Pasting the bank statement
The common mistake is uploading a full statement so the agent can categorize it. Upload the receipt images only, and keep account numbers out of the file.
How to choose and combine
Once you know what each agent finishes, the remaining work is trust and stacking. Which answer do you believe, and what do you add when the big four stall on your particular chore?
These three entries settle that, then the close picks by chore.
8. All four plan the same trip to the same places
Test every agent on one three-day trip and the itineraries come back near-identical. WIRED ran a China trip through each agent it tested and got the same routes from all of them.
That is the Lonely Planet effect: the crowd gets pushed to the same places.
The bigger problem is the answer itself. Every one of the four can state a false fact with confidence and invent a source, a quote or a link. A route that looks researched can rest on a page that never existed.
Here is the check. Paste the same trip request into a second agent and compare the two lists side by side. Ask for the source list before the answer, and open each link before it reaches a booking.
Prompt to copy
Here is my three-day trip: [city, dates, budget, who is traveling]. Give me a day-by-day plan with each stop, the travel time between stops, and a source link for every opening hours or price you state. Mark anything you are not sure of as unverified, and do not fill gaps with guesses.
The run costs 15 minutes for the trip and 20 minutes of review, which includes opening the links. Five minutes of comparison across two agents is the cheapest part of the job.
Two-agent check: 15 minutes to run, 5 minutes to compare, 20 to review. The invented link gets caught before it turns into a paid reservation.
The Trust Traffic Light
- 🛒 Agent builds a cart for you to review — Safe as long as you open it before paying.
- 📄 Agent drafts an expense sheet from receipts — Check the totals against the images.
- 💳 Agent uses a one-time card number — Fine with a spending limit set first.
- 🔗 Agent supplies a source link you have not checked — Verify before quoting it anywhere.
- 🏦 Agent logs into your bank with your password — Never. Generate a statement extract instead.
- 📞 Agent calls a clinic for an appointment — These tools struggle with calls. Do it yourself.
Based on PCMag and WIRED agent testing and Forbes India user reports.
9. When the big four stall, add a specialist
Prompt to copy
I manage email, scheduling and client reminders for a one-person consulting business, handling about 40 emails a day. Compare Lindy, Zapier Agents and Instinct for automating my inbox triage and meeting prep. Use one table: tool, task it finishes, monthly cost from the vendor's page, team or solo fit, and setup effort. Flag any claim you cannot source. Under 300 words.
The catch: none of these three replaces the big four, and each charges for a narrow slice. Zapier Agents lists 7,000+ app integrations on a freemium plan from about $20/mo, per Oxylabs, while your bill climbs with every task it runs.
Lindy goes after inbox sorting, meeting prep and reminders through your work tools and texts, per Zapier, but it suits solo work over teams and costs more per user. Instinct stays invite-only, answers over iMessage, WhatsApp or a call, and reads Hindi voice notes.
📅 The First Week With One Agent
Pick one repeating chore
Choose the chore you do weekly that ends in a cart, a file or a calendar invite.
Turn off model training
Open settings on the free tier and switch off the option to use chats for model improvement.
Run the chore once
Send the task and let the agent finish, noting where it stops and asks for a login.
Time the review
Clock the minutes you spend checking and editing, because that is the real cost.
Compare against by hand
Put your review minutes beside the by-hand time for the same chore.
Keep or cancel
Pay for the agent only if the total beats the old way and the chore repeats.
10. Pick by the chore you hate most
One agent per chore, then one signup. Dots is the pick for shopping and job boards. Spark is the pick for research and trip plans.
Muse is the free pick for general use in the US and Canada. Claude Duet is the pick for work admin and expense sheets.
If $100 a month is too much, Dots is out. Spark at $20 is the next best overall. For a grocery cart, the free option is Muse, and you log in yourself and watch for wrong items.
The setup takes about ten minutes, once. Open one free tier. Run your one repeating chore one time. Time the review and edits that follow.
Frequently Asked Questions
Do I need to know how to code to use these agents?
No. All four work through a chat box or an app, and you describe the job in plain words. Cowork on macOS asks you to point it at a folder, which is the closest thing to a technical step.
Is the free tier of Muse enough to finish a grocery order?
It can finish, but expect manual Instacart logins and a slow run, plus the risk of a wrong item added to the cart. Review every line before you pay.
Why is Dots only on the $100 plan?
According to PCMag Australia, Dots is included in ChatGPT for Pro, Business and Enterprise subscribers. The free tier does not include the agent feature, so the $100 is the entry point.
Can these agents schedule a doctor's appointment for me?
Not reliably. Users report they struggle with phone calls and appointments, per Forbes India. Use the agent for research and paperwork, then make the call yourself.
What should I never paste into an agent?
Passwords, client files, and medical or financial records. Describe the task instead of uploading the underlying document, and switch off model training in settings on any free plan.
Which Agent Gets Your Card This Month
Pick by chore: Dots for shopping and job boards, Spark for research and trip planning, Muse for free general use in the US and Canada, Claude Duet for work admin and expense sheets.
If $100 a month is too much, Spark at $20 is the runner-up.
This week, pick the one chore that repeats, run it on a single free tier, and time the review rather than the agent. That number tells you whether a subscription earns its place.