Best AI agent is the title every big AI company now wants, and WIRED has just run one of the first head-to-head tests to settle it. On 27 October 2026, WIRED software reporter Reece Rogers published the results of a 24-hour experiment. He set up six always-on agents on his iPhone: Meta’s Muse, OpenAI’s Dots, Instinct, Anthropic’s Claude Duet, Google’s Gemini Spark and xAI’s Grok Bot. He then put each one through the same five everyday tasks and scored every attempt from 1 to 5.
The winner was OpenAI’s Dots, the most expensive agent in the test at $100 a month. Meta’s Muse, which is free, came a close second. Claude Duet, which had launched as a research preview only the night before, won praise for the best interface. Instinct took the award nobody wants: most annoying.
This guide walks through WIRED’s test in detail. It covers the six contenders and their prices, the five tasks, the scores for each one, the awards, and the privacy catch that comes with every agent. We have also added up WIRED’s individual scores into a single table, so you can see exactly how the best AI agent pulled ahead. For background on the two leading agents, see our earlier piece on whether you can trust Meta’s Muse or OpenAI’s Dots to run your life.
Table of contents
- Who Won the Best AI Agent Test?
- The Six Contenders in the Best AI Agent Test
- How WIRED Judged the Best AI Agent
- Test 1: Which Is the Best AI Agent for News Briefings?
- Test 2: The Best AI Agent for Ordering Lunch
- Test 3: Can the Best AI Agent Catch a Schedule Clash?
- Test 4: The Best AI Agent for Planning a Trip
- Test 5: The Best AI Agent at 3 am
- The Best AI Agent Awards
- Strengths and Weaknesses at a Glance
- Privacy and Security: The Catch With Every AI Agent
- How to Choose the Best AI Agent for You
- Our Verdict on the Best AI Agent
- Best AI Agent FAQs
- References
Who Won the Best AI Agent Test?
The short answer is OpenAI’s Dots. Rogers named it “Best Overall” and wrote that it “performed the best overall in these tests”. He also noted that it failed at one task, the news pitch, which he described as “fairly specific to my niche”. Everywhere else, Dots scored 4 or 5.
The final scores
WIRED published a score for every agent in every test, but did not print a final total. We added them up. Each agent could score a maximum of 25 points across the five tests.
Bars are scaled to the top score: 18 is 90% of 20, 16 is 80% and 12 is 60%. The totals match WIRED’s own description. Rogers wrote that Muse came “in a close second on the scorecard” and that Claude Duet scored “in the middle of the pack on most tests”.
How the totals add up
The table below shows every score WIRED gave, test by test. It makes clear why Dots is the best AI agent in this test even though it lost the first round badly.
| Agent | News pitch | Lunch order | Schedule clash | Trip plan | Insomnia | Total |
|---|---|---|---|---|---|---|
| Dots | 2 | 5 | 5 | 4 | 4 | 20 |
| Muse | 3 | 4 | 2 | 5 | 4 | 18 |
| Grok Bot | 1 | 4 | 2 | 4 | 5 | 16 |
| Claude Duet | 4 | 3 | 4 | 3 | 2 | 16 |
| Gemini Spark | 5 | 1 | 1 | 3 | 2 | 12 |
| Instinct | 1 | 2 | 2 | 3 | 4 | 12 |
Two patterns stand out. Dots never dropped below 4 after the first test, which is what made it the most reliable agent. Gemini Spark had the opposite profile: it won the news pitch outright with a perfect 5, then scored 1 on both the lunch order and the schedule clash.
The Six Contenders in the Best AI Agent Test
All six agents competing for the best AI agent title work in a similar way. They use natural language processing to understand your messages. You message them, they act on your behalf, and they can message you first when they spot something useful. But they differ a great deal in price, in where they live on your phone, and in how easy they are to get.
| Agent | Maker | Price in the test | Where it lives | Access |
|---|---|---|---|---|
| Muse | Meta | Free | Its own app | Open since its September release |
| Dots | OpenAI | $100 a month | Inside ChatGPT, plus a widget | Paid ChatGPT plan |
| Instinct | Instinct | Free | Your text messages | Invite code from an existing user |
| Claude Duet | Anthropic | $20 a month | A dedicated app | Wait-listed research preview |
| Gemini Spark | $20 a month | Inside the Gemini app | Beta | |
| Grok Bot | xAI | $30 a month | Its own app | Paid plan |
The price spread is large. The two free agents, Muse and Instinct, sit at one end. Dots sits at the other, at five times the price of Claude Duet or Gemini Spark.
Bars are scaled to the Dots price: $30 is 30% of $100 and $20 is 20%. The free agents are drawn as a sliver so they remain visible. Put next to the score chart, it shows the trade-off clearly: Muse gets 18 of Dots’ 20 points for none of the cost.
Why OpenClaw sat this one out
One well-known name is missing. OpenClaw, the open agent that sparked much of this year’s interest, was left out because Rogers judged it “too clunky for the average person to set up”. He said a future ranking could include it, along with smaller players such as Hermes by Nous Research and Poke. So this is a test of the best AI agent for ordinary phone users, not for developers.
How WIRED Judged the Best AI Agent
Rogers has tested AI software “from a nontechnical perspective for the last few years”, and he designed the test around that view. He set up all six agents before bed on a Wednesday, then ran the tasks over the following 24 hours.
Five tests in 24 hours
The five tests were chosen to measure three things: how proactive each agent was, how well it followed directions, and how well it could browse the web. Each attempt was scored from 1 to 5, and the totals decided the awards. The tests were:
- A morning news briefing that ended in a story pitch for WIRED’s weekly meeting.
- A takeaway lunch order for Rogers and his vegan boss.
- A hidden clash between a calendar meeting and an emailed reminder.
- A first draft of a two-week trip to China.
- An open-ended request for help in the middle of a sleepless night.
What the test could not measure
Rogers is clear that this is “not an endorsement of any of these AI agents“. He connected only his personal Google Calendar and Gmail, so the agents had far less personal data than a heavy user would give them. All of them should improve as they learn more about a user. A one-day test also cannot show how reliable an agent is over weeks. Treat the result as a useful snapshot of the best AI agent today, not a final verdict.
Test 1: Which Is the Best AI Agent for News Briefings?
Companies often sell agents as a way to build a morning briefing. Rogers sent each agent a long prompt describing his beat at WIRED and asked it to spend 30 minutes researching before suggesting a story for the Gear desk’s Thursday meeting.
Who followed the instructions
Only Muse and Dots kept researching for most of the 30 minutes. The other four either replied at once or sent a headline after a couple of minutes. Rogers called this “not a major mistake”, but said it suggests Muse and Dots “were a bit better at following my instructions—a critical function of an AI agent”. If you are looking for the best AI agent for long research jobs, that detail matters.
Gemini Spark’s winning pitch
The quality of the pitches was a different matter, and here the best AI agent was not Dots. Gemini Spark scored a perfect 5 for a story contrasting Walmart, which has built ways for outside agents to shop on its platform, with Amazon, which “recently gave Muse the boot”. Claude Duet scored 4 for a pitch about ChatGPT’s new PIN-locked chats. Muse scored 3, Dots 2 and both Grok Bot and Instinct 1. Grok Bot’s idea was “old news”, and Instinct’s was “bland and hard-to-measure”.
Rogers admits the pitches “didn’t exactly go over well” with his colleagues. His team, he wrote, “didn’t seem too concerned about getting scooped by an agent”.
Test 2: The Best AI Agent for Ordering Lunch
The second test was about small daily purchases. Rogers asked the agents to plan a takeaway lunch near the WIRED office. His boss is vegan and wanted Vietnamese food. Rogers wanted something meaty with plenty of caffeine. He promised to order whatever the best AI agent picked.
| Agent | Restaurant | Ordering route | Cart subtotal | Score |
|---|---|---|---|---|
| Dots | DragonEats | Restaurant’s own website | $30.57 | 5 |
| Muse | DragonEats | DoorDash | $42.66 (items) | 4 |
| Grok Bot | Freshroll | DoorDash | $43.40 | 4 |
| Claude Duet | Freshroll | Asked: DoorDash or Uber Eats | $49.25 | 3 |
| Instinct | Freshroll | DoorDash | $53.73 | 2 |
| Gemini Spark | Freshroll | Pickup | $22.55 | 1 |
All subtotals are before tax and fees, as quoted by the agents in WIRED’s article. Muse’s figure is an item total, because DoorDash does not show fees until checkout.
Dots built the order Rogers placed
Dots won this round, and Rogers actually ordered what it put in the cart: a roast pork bánh mì and a Vietnamese iced coffee for him, and teriyaki tofu fresh rolls with soy sauce instead of fish sauce for his boss. It was also the cheapest complete order at $30.57. Muse built a very similar cart at DragonEats through DoorDash. Instinct and Claude Duet both chose pricier carts at Freshroll, with Instinct’s about $4 more than Claude Duet’s.
Gemini Spark’s snack-sized lunch
Gemini Spark was the only real failure, and the furthest from the best AI agent in this round. It ordered a single spring roll for the vegan boss, “more of a snack than anything”, in Rogers’ words. Its low subtotal of $22.55 came from leaving most of the boss’s lunch out. Cheap is not the same as correct, and that is a useful lesson for anyone handing purchases to an agent.
Test 3: Can the Best AI Agent Catch a Schedule Clash?
Agents are sold as proactive, so Rogers set a trap. He created a Google Calendar event at 2 pm, labelled as an important source meeting that could not be moved. He then emailed himself a made-up reminder that his mother was leaving hospital and he needed to call her at 2 pm. The question was simple: which agent would spot the clash and warn him before it happened?
Only two agents spotted the clash
Only two did, which was fewer than Rogers expected. Dots scored 5: it “flagged the issue almost immediately” and suggested calling before or after the meeting. Claude Duet scored 4 for flagging the conflict hours ahead, though it did not suggest an alternative.
The other four missed it. Grok Bot, Instinct and Muse each scored 2, because they at least offered to check emails against the calendar in future. Gemini Spark scored 1. It neither flagged the clash nor offered a fix. Its reply explained that the reminder “arrived as an email message rather than a calendar entry”, so the overlap “wasn’t surfaced”.
Why this test matters most
For many people this is the task that decides the best AI agent. A briefing or a trip plan can be checked later. A missed clash cannot. The test shows that most agents still treat each connected app separately, and that joining information across email and calendar remains hard. It also shows why Dots and Claude Duet stood out as the most proactive helpers in the group.
Test 4: The Best AI Agent for Planning a Trip
For the fourth test, Rogers asked for a first-draft itinerary for a two-week trip to China late next year with his partner. He wanted flights, hotels, nightlife and food tips, a mix of history and big-city life, and no more than four hotels.
Every agent picked the same route
The most striking result had nothing to do with scores. All six agents chose the same route: Beijing, Xi’an, Chengdu and Shanghai. Rogers warned that this could create a new “Lonely Planet effect”, where many travellers are steered to the same places, shops and restaurants. It is a reminder that agents built on similar data tend to give similar answers.
Muse made the most readable plan
Muse was the best AI agent for travel planning, scoring 5 for an itinerary that was “extremely easy to read, with visual tables and straightforward tips”. Dots and Grok Bot scored 4. Dots’ plan was strong but ran to 24 pages, which Rogers found excessive. Gemini Spark, Instinct and Claude Duet each scored 3. Claude Duet used no images or visual elements, though Rogers still called its plan “solid”.
Rogers also added a sensible warning. He recommends checking non-AI sources before trusting any itinerary, and points to a group in California that had to be rescued after relying on a chatbot to plan a demanding hike.
Test 5: The Best AI Agent at 3 am
The final test was the most open-ended. Rogers could not sleep, which he partly blamed on the Vietnamese coffee, and he asked the agents for help. The goal was to see how each agent responded to a vague request with no clear task.
Grok Bot’s best moment
Grok Bot was the best AI agent in this round, with a perfect 5. It offered to set a reminder to check on Rogers’ mother later that morning “so it’s off your mind for now”. Rogers said that if the scenario were real, he “would likely follow through on it”. Dots, Instinct and Muse each scored 4. Dots asked him to send “the three things taking up space in your head” so it could turn them into a checklist.
Where Claude Duet and Gemini Spark slipped
Gemini Spark and Claude Duet both scored 2, and Rogers said they “really flopped”. Gemini Spark’s advice was fine, but it pointed out that his morning calendar was empty, which he found annoying. Claude Duet offered to close out “yesterday’s two loose ends”, the news pitch and the lunch order. Rogers called it one of Claude Duet’s weakest answers, because neither task was on his mind at 3 am.
The Best AI Agent Awards
Rather than rank all six agents in order, WIRED handed out four awards. Here is what each one means for anyone choosing the best AI agent.
Best overall: OpenAI Dots
Dots was the best AI agent overall. Rogers said that in the month since its release “Dots already feels more polished”. His main complaint is price. He suggests most people wait until OpenAI offers Dots for free, or at least on the $20-a-month plan. He would also prefer a separate app to the current ChatGPT integration, though he called the smartphone widget “a decent compromise”. For more on how Dots grew out of OpenAI’s coding tools, read our piece on how OpenAI turned its coding tools into Dots.
Most accessible: Meta Muse
“Muse is for the normies! And that’s fine,” Rogers wrote. It came a close second on the scorecard, was the easiest to set up and is free. His concern is privacy. Meta says it processes user information in secure virtual machines, but Rogers still has doubts “given the company’s past mishandling of user data”. For many people, Muse may still be the best AI agent for the money.
Best interface: Anthropic Claude Duet
Claude Duet scored in the middle of the pack, but its dedicated app was “extremely user-friendly” for following up on ongoing tasks. Rogers tested it one day after it launched as a research preview, so he expected rough edges. He thinks productivity fans could take to it quickly. For now, you need to join Anthropic’s waiting list and pay $20 a month.
Most annoying: Instinct
Instinct works entirely through text messages, and Rogers did not like it. Other WIRED writers have praised the agent, but he found that the texting-only design left him “too removed from what my agent was doing”. He was also irritated by its emoji reactions, including a grimace face when he pointed out it had missed his schedule clash. Our earlier comparison of Instinct and Muse covers why the agent built such a following.
Where Grok Bot and Gemini Spark landed
Neither Grok Bot nor Gemini Spark won a best AI agent award. By our count, Grok Bot tied Claude Duet on 16 points, helped by its perfect insomnia answer and strong lunch and travel plans. Gemini Spark tied Instinct on 12. It produced the best news pitch of all, but its failures on lunch and scheduling cost it heavily.
Strengths and Weaknesses at a Glance
The table below sums up each agent’s best and worst moments in the best AI agent test, using WIRED’s scores and comments.
| Agent | Best moment | Worst moment | Best for |
|---|---|---|---|
| Dots | Spotted the 2 pm clash at once (5) | An evergreen news pitch (2) | Reliability across tasks |
| Muse | The clearest trip plan (5) | Missed the clash (2) | Free, easy everyday use |
| Grok Bot | Kind, practical 3 am advice (5) | An old-news pitch (1) | Readable plans and tips |
| Claude Duet | A timely pitch and an early clash warning (4) | Tone-deaf 3 am reply (2) | A clean app for tracking tasks |
| Gemini Spark | The best news pitch (5) | A one-roll lunch and a missed clash (1) | Ideas and angles |
| Instinct | A sensible bedtime task (4) | A bland pitch (1) | People who live in their texts |
No agent was best at everything. That is the real lesson of the test: the best AI agent depends on which jobs you plan to give it.
Privacy and Security: The Catch With Every AI Agent
Rogers stresses that agents “work best when connected to your personal data, like email and calendars”, and that this requires a level of trust some people will not want to give. Even the best AI agent is only as safe as the access you give it. The software, he writes, “remains nascent and often fraught with privacy and security concerns”.
The Slack bank balance story
He gives one example. An X user claimed he had connected AI agents to his bank account and his work Slack, then was “mortified” when an agent posted how much money he had where his colleagues could see it. It is a single anecdote, but it shows how quickly an agent with broad access can share the wrong thing in the wrong place.
What to connect, and what to keep back
Rogers limited his own risk by connecting only his personal Google Calendar and Gmail. That is a sensible model for anyone trying out the best AI agent for the first time. Start with low-risk accounts, avoid connecting financial or work systems, and watch what the agent does before you widen its access. Our guide to the privacy promises AI agent makers are making explains what each company says it does with your data.
How to Choose the Best AI Agent for You
WIRED’s test gives a clear winner, but the right choice depends on your budget and your needs. Here is how we would read the results.
If you want the best AI agent overall
Choose Dots if you already pay for a top ChatGPT plan or can justify $100 a month. It was the most consistent agent in the test and the best at the proactive work that agents are meant to do. If you cannot justify the cost, Rogers’ advice is to wait for a cheaper tier.
If you want the best AI agent for free
Muse is the obvious pick. It scored 18 of a possible 25, close to Dots, and costs nothing. Weigh that against your view of Meta’s handling of personal data. Instinct is also free, but it scored lower and only works through text messages.
If you care most about the app
Claude Duet had the best interface in the test and a strong record on spotting the schedule clash. It is still a waiting-list preview, so expect it to change quickly over the next few months.
For businesses testing agents
For a business, the best AI agent is the one that follows instructions and handles data safely, not the one with the cleverest ideas. Run your own small test, much like WIRED’s: a few real tasks, a simple score for each, and a limited set of connected accounts. Our team can help you design that kind of pilot through our AI employees and autonomous agents service, and set guardrails as part of a wider AI strategy.
Our Verdict on the Best AI Agent
WIRED’s test is a snapshot, but a useful one. A year ago, Rogers was complaining that AI agents “suck”. Now he writes that they work, “at least more reliably than last year”, and that “these little guys got their glow-up”.
On this evidence, OpenAI’s Dots is the best AI agent you can buy today, and Meta’s Muse is the best one you can use for free. The gap between them is two points, while the gap in price is $100 a month. Claude Duet is the one to watch as it leaves preview. Whichever you choose, connect as little personal data as you can, check its work, and remember that every one of these agents can still miss something obvious.
Best AI Agent FAQs
Which is the best AI agent in WIRED’s test?
OpenAI’s Dots. It won WIRED’s “Best Overall” award, and by our count of WIRED’s scores it finished with 20 points out of 25, two ahead of Meta’s Muse.
Which AI agents did WIRED test?
Six: Meta’s Muse, OpenAI’s Dots, Instinct, Anthropic’s Claude Duet, Google’s Gemini Spark and xAI’s Grok Bot. OpenClaw was left out because it is hard for most people to set up.
What is the best free AI agent?
In this test it was Meta’s Muse, which scored 18 points and won “Most Accessible”. Instinct is also free but needs an invite code and scored 12.
How much does the best AI agent cost?
As tested, Dots cost $100 a month, Grok Bot $30, Claude Duet and Gemini Spark $20 each, and Muse and Instinct were free.
Is it safe to connect an AI agent to my email?
It carries real risk. WIRED connected only a personal calendar and Gmail account. Start small, avoid financial and work accounts, and check what the agent does before giving it more access.
What is the difference between an AI agent and a chatbot?
A chatbot waits for your prompt and replies. An agent works more like a digital assistant: it can act on your behalf, often by controlling a virtual browser, and it can message you first when it spots something useful.
References
I Tested Every Major AI Agent for a Day. Here’s the Best One (WIRED)
Get started with Claude Duet (Anthropic support)
The next evolution of the Gemini app (Google)
Meta releases Muse, a personal AI agent (WIRED)
OpenAI’s Dots are always-on AI agents (WIRED)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
