AI agents spent a long time as that shiny promise that looked great in a demo and stumbled in real life. Lots of slick showcases, very few tasks actually done. Reece Rogers, a reporter at WIRED, says that last year he was always complaining to his newsroom colleagues about how bad these tools were at almost everything. According to the outlet, things have changed. The agents still make mistakes, but now they finish tasks and handle the browser with a much steadier hand than before.
To find out who is delivering the most, Rogers ran a very down-to-earth AI agent test. He installed six agents on the same iPhone and spent about 24 hours putting all of them to work on simple tasks. The lineup includes Muse, from Meta, and Claude Duet, from Anthropic.
Below, you will see how the experiment was set up, what the first round of tests revealed and which warnings the reporter leaves along the way. You will also learn how to choose an AI agent for your business in Brazil. Do not expect a ready-made ranking to copy. The goal is to walk away with criteria that make sense for your own routine.
What an AI agent is, without the jargon
Think of a digital assistant you reach by message, as if you were texting someone on your team. You ask, and it does the work. That could mean digging through your inbox for something important you missed, keeping an eye on a flight price or helping out with an everyday buying decision.
The big difference from a regular chatbot is initiative. With a chatbot, the conversation happens in turns: you ask, it answers, and that is it. An agent goes further. When it spots a chance to help, it starts the conversation on its own.
Picture, just as an example, someone with a meeting on the calendar and an email from a client asking to move the time. A chatbot would only know about it if the person asked. An agent connected to email and calendar can notice the conflict and flag it ahead of time. That is the kind of behavior companies are selling.
There is one more detail that explains why these tools got more useful. According to WIRED, many agents work with their own virtual browser, almost a computer all to themselves, to visit websites and take action online. And conversations usually live in a single thread that looks a lot like a group chat.
The six agents in the test
According to the WIRED report, the group included Muse, from Meta, Dots, from OpenAI, Instinct, Claude Duet, from Anthropic, Gemini Spark, from Google, and Grok Bot, from xAI. Each one has its own way of working and its own pricing. Instinct is free, but you can only get in with an invite code from another user, and everything happens through text messages, with no app. Gemini Spark is in beta inside the Gemini app and costs US$ 20 a month. Grok Bot runs US$ 30 a month and has its own app. Meanwhile, Claude Duet, the new agent app from Anthropic, costs US$ 20 a month and was the newest of the bunch: it launched as a research preview, with a waitlist, the night before the tests.
The price extremes stand out. Muse is free and, according to WIRED, has topped the mobile app download charts since it launched in September. At the other end is Dots, which works inside ChatGPT and costs US$ 100 a month, the highest price among the tools reviewed.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Some names were left out, too. OpenClaw, which the outlet says helped drive much of this wave of innovation this year, did not make the list because it is too complicated for an ordinary person to set up. Rogers also mentions smaller names, like Hermes, from Nous Research, and Poke, as candidates for a future comparison.
For readers in Brazil, a quick heads-up. All prices are in US dollars, so the final cost changes with the exchange rate and the taxes on international card purchases. And not everything that launches abroad arrives here at the same time. Before you subscribe, check whether the agent is available in the country and whether it works well in Portuguese.
How WIRED set up the test
The scoring was simple. And that is great news for anyone who wants to try the same idea in their own business.
Rogers created five basic tasks, gave each attempt a score from 1 to 5 and added everything up at the end to pick a winner. As a bonus, he promised an award for the most annoying agent, which already says a lot about how pushy these tools can be.
Three skills were at the heart of the evaluation:
- Initiative: whether the agent acts on its own when it makes sense;
- Following instructions: whether it does what was asked, the way it was asked;
- Navigation: whether it can find its way around the internet on its own.
The reporter himself admits an important limit. He connected only his personal Gmail and Google Calendar, without opening up the rest of his digital life to the tools. Since these agents improve as they get more data and learn the habits of the person using them, the results show only a slice of what they can do.
The first result: the morning briefing
The first task came from a real situation. Every Thursday morning, the consumer tech team Rogers works on meets to discuss story ideas. He asked each agent for a roundup of tech news and a story pitch to bring to the meeting, with plenty of context about the topics he covers and how the idea would be used.
Picking this test makes sense. Companies often sell agents as perfect for putting together that morning digest that keeps you up to speed on what matters.
Who followed orders and who had the best idea
The request had a catch: spend 30 minutes researching in depth before answering. According to WIRED, only Muse and Dots respected that and spent most of the time researching. The others sent back a pitch right away or after just a few minutes of searching. It is not a huge mistake, but it shows who took the instruction seriously.
Following orders, however, did not guarantee the best idea. Gemini Spark earned a perfect score with a pitch contrasting Walmart, which created ways for third-party agents to use its shopping platform, with Amazon, which had just made headlines for blocking Muse, the Meta AI agent, from its e-commerce site. Claude Duet came right behind with a 4. Muse got a 3. Dots got a 2 for pitching a topic that felt more evergreen than timely. Grok Bot came in last with a 1 because it brought up old news.
And the meeting? Rogers says, with good humor, that the pitches did not excite his colleagues. Nobody in the newsroom seemed worried about losing their seat to an agent anytime soon.
For business owners, the lesson is straightforward. Following a request and delivering something good are two different skills. The best AI agent for your daily work needs both, and in this round no tool shined on both fronts at once.
The winning pitch, by the way, touches a sensitive point for retail. The fight between retailers that open their doors to agents and retailers that shut them out directly affects anyone who sells online. If agents start shopping on behalf of people, the way a store shows up to these tools could matter as much as the way it shows up to human customers.
The warning nobody should skip: privacy
WIRED makes a point of saying the test is not a buying recommendation. The technology is still early and comes with serious privacy and security questions. These agents work best when they can access your data, like email and calendar, and that takes a level of trust that may scare off a lot of people.
The outlet mentions a story that works as a cautionary tale. One user on X said he connected agents to his bank account and his work Slack. One of them ended up posting how much money he had, right where his coworkers could see it. Embarrassing, to say the least.
Now bring that inside a company. Imagine a small agency that connects an agent to a shared inbox and the team chat to speed up replies. All it takes is one misunderstanding for a sales proposal or client data to land in the wrong place. Customer data, team information and financial numbers call for extra care, especially with LGPD, the Brazilian data protection law, on the radar. If you want to dig deeper, you can read about AI, security and autonomous agents in business, a topic that is only going to grow.
How to choose an AI agent for your business in Brazil
Start with the task, not the tool. Before looking at any ranking, define what you want to solve. It could be a daily briefing, customer support help, market research or that admin routine nobody can stand anymore. The first WIRED round already showed that the most obedient agent was not the most creative one, and vice versa.
Next, run your own small test. Pick three or four real tasks from your operation and give simple scores, just like Rogers did. Picture an online store that has two agents summarize the customer questions of the week and compares which summary helps the team more. Or a marketing team that asks each agent for a list of blog topics and checks which ideas are actually fresh. In a few days, you can feel the difference.
Also pay attention to how each tool behaves when nobody is watching. Does it stick to the deadline you set? Does it send too many messages? Does it make things up when it cannot find an answer? These little quirks weigh a lot in daily use.
At first, connect as little data as possible and grant access gradually as trust grows. Factor in the cost in dollars and the time your team will spend reviewing what the agent delivers. A cheap tool that needs constant review can end up costing more than it seems.
And remember that AI agents for business are still maturing. They help a lot, but they do not work alone. Human review is still a must, especially for anything involving customers, money or sensitive data.
What the test leaves behind
AI agents have improved, and the WIRED experiment makes that clear. But improving does not mean any of them solves everything. In the briefing round, one researched more patiently, another had the most interesting idea, and a third still tripped over the basics by bringing up old news.
That is why the best choice depends on the task that weighs most on your routine. Someone else’s ranking is a starting point, not a final answer. What worked for a tech reporter in the United States may not work for a marketing agency in Belo Horizonte or an online store in a small town.
Here at Método Viral, our coverage of AI agents keeps a close eye on every launch and every limit that shows up in practice. If you have already tried one of these agents, tell us in the comments which one it was and how it did in your work.
Set aside one boring task from your week, pick two agents and run your own test with a simple scorecard. With little data connected and a sharp eye on the results, you will quickly find out which tool deserves a spot on your team.
