WIRED’s 24-Hour AI Agent Test Finds Meta’s Muse Leads a Sharper, Still-Fallible Pack

WIRED’s 24-hour test of six always-on AI agents on an iPhone found that Meta’s Muse was the strongest overall performer, while Google’s Gemini Spark handled one briefing task best and OpenAI’s Dots did the best job spotting a conflict before it became a problem.

The results suggest that consumer AI agents are getting more useful at research, scheduling and follow-through, but they still miss basics, make strange choices and vary sharply in value for money.

What WIRED tested

WIRED software reporter Reece Rogers set up six always-on AI agents on an iPhone on Wednesday night and used them through Thursday in a series of everyday tasks. The lineup included Meta Muse, OpenAI Dots, Instinct, Anthropic Claude Duet, Google Gemini Spark and xAI Grok Bot.

The test was designed to see how these products behave when they are expected to stay active, keep researching and help with mundane planning rather than just answer a one-off prompt. WIRED said the agents are improving at task follow-through and browser control, but still make mistakes.

How were the agents judged?

WIRED used a small set of practical tasks: a briefing assignment, lunch planning, travel suggestions and a calendar-conflict test. The reporter also examined how long each product kept working, how confidently it handled instructions and whether it surfaced important details on its own.

That made the ranking less about abstract benchmark scores and more about whether an agent could actually be useful in a normal workday.

Who came out on top?

Meta’s Muse was WIRED’s overall winner. The product performed best across the range of tasks and was the only agent the reporter judged strong enough to place first overall.

Gemini Spark earned the top mark on one specific assignment: it produced the strongest morning-news angle in the briefing exercise. OpenAI’s Dots stood out for conflict detection, identifying the email-calendar problem before the overlap happened.

“They were just bad at everything.”

That line from Reece Rogers reflects how much the category has moved in a short period. WIRED said last year’s agents were mostly disappointing; this year’s group was markedly more capable, even if none of them was perfect.

By the numbers

The most revealing part of the test was not just which company won, but how often the products converged on the same answers and where they still broke down.

What did the agents get right — and wrong?

The briefing task was one of the clearest tests of persistence. Only Muse and Dots kept researching for most of the allotted 30 minutes, which gave them an edge over the rest of the field.

In the morning-news angle exercise, Gemini Spark produced the strongest result. But the same product later stumbled on lunch planning, adding only one $5 roll for the boss, which was not a useful real-world recommendation.

On the calendar-conflict challenge, only two agents flagged the problem before the overlap. That matters because scheduling is one of the most appealing use cases for personal AI agents, and the test showed that even when the systems can access personal data, they do not always interpret it safely or completely.

Rogers also found that Muse and Dots both asked whether to call mom before or after the meeting, an example of how some agents are starting to show a semblance of practical planning around daily life rather than just generating text.

Where the tools converged

Several of the agents chose DragonEats or Freshroll for lunch, suggesting that when the models are asked for convenience-driven recommendations, they tend to cluster around similar mainstream options.

The same pattern showed up in travel planning. The agents offered similar China-trip routes, which points to a broader tendency in AI-generated travel and shopping advice: if multiple models are optimizing for obvious, popular choices, they can end up steering users toward the same destinations and itineraries.

Why it matters

The test matters because AI agents are moving from novelty into personal utility. These products are not just chatbots that answer questions; they are designed to act more proactively and use tools such as virtual browsers.

That shift changes what users expect. If an agent can monitor email, scan calendars, plan meals and draft itineraries, it can save time in concrete ways. But the WIRED test shows that those same strengths create new risks: a missed conflict, a sloppy booking choice or a badly judged recommendation can cause real inconvenience.

Privacy is part of the equation too. These agents often connect to personal data such as inboxes and calendars, which means the convenience comes with security concerns. The more helpful the software becomes, the more sensitive information it may need to touch.

The pricing differences also matter. A product like OpenAI Dots, at $100 per month, has to justify a premium that is five times the cost of Gemini Spark or Claude Duet and more than three times the cost of Grok Bot. Meta’s Muse being free gives it an obvious consumer advantage if its performance remains strong.

How did the products compare on price and access?

The clearest split in this test was not only performance but access model. Some products are free, some require subscriptions, and one is still gated by an invite code.

Instinct’s setup proved too clunky for average users, which is why WIRED excluded OpenClaw from the test. That detail underscores a recurring point in consumer AI: a tool can sound promising on paper and still fail because the setup is too cumbersome for ordinary people.

Background: how we got here

AI agents have been pitched as the next step beyond chatbots. Instead of merely responding to a query, they are meant to take action, remember context and work through connected tools on behalf of the user. In theory, that makes them more useful for real-life errands such as planning travel, triaging messages and managing calendars.

In practice, the category has taken time to mature. WIRED notes that last year’s agents were mostly bad at their jobs. The current crop appears more polished, which is why this test is important: it captures a moment when the products are finally becoming usable enough to compare in a meaningful way.

Meta’s Muse has gained traction quickly. WIRED said it has topped smartphone app download charts since its September release, a sign that distribution and consumer familiarity may matter nearly as much as raw capability.

Anthropic’s Claude Duet entered the test in a rushed state, arriving as a wait-listed research preview the night before. That made it a late entrant in a field where fast-moving product launches can change the picture almost overnight.

What this says about the market

For consumers, the message is mixed but encouraging. There are now products that can do parts of the job well enough to be genuinely helpful. Muse, Gemini Spark and Dots each demonstrated a distinct strength, suggesting the market is moving toward specialization rather than a single dominant model that excels at everything.

For vendors, the test is a reminder that premium pricing will face pressure unless the product is consistently reliable. A high subscription fee is harder to defend if a free alternative can perform better overall.

For the broader market, the test also hints at a coming problem of convergence. When multiple agents recommend the same lunch spots or travel routes, the value of personalization starts to erode. Users may get speed, but not necessarily variety or judgment.

Why the privacy question is getting sharper

Personal AI agents rely on access to sensitive data because that is what makes them useful. Email and calendars are the obvious entry points, but those same integrations create the possibility of mistakes that are more consequential than a wrong answer in a chatbot window.

WIRED’s conflict-detection test is therefore a useful warning. Only two agents spotted the issue before the overlap. That is a promising result compared with earlier generations of software, but it is not good enough to hand over important scheduling decisions without supervision.

What happens next

The next question is whether any of these agents can prove themselves in broader real-world use beyond a single reporter’s one-day test. The ranking offers a useful snapshot, but it does not settle which system will work best for different kinds of users or tasks.

Another open issue is reliability in sensitive workflows. Inbox triage and scheduling are especially tempting uses because they directly save time, but they are also the places where mistakes can create the most friction. The test suggests that the software is getting closer, though not close enough to run unattended.

Pricing will also be worth watching. If Muse keeps delivering solid results while remaining free, it could put pressure on competitors that charge monthly fees. But if its advantage narrows, the market may fragment into a tiered landscape where each service wins only in a narrow slice of tasks.

Finally, the convergence in travel advice is a sign that AI recommendations may become more standardized as they improve. If more travelers use the same agents, they may end up following the same routes and booking the same places, which could reshape how people discover destinations.

Bottom line

WIRED’s test suggests that AI agents are no longer merely a promise. They are useful enough to compare seriously, and in some cases useful enough to trust with specific jobs.

But the report also makes clear that the category is still uneven. Meta’s Muse led overall, Gemini Spark shone on a briefing assignment, and OpenAI’s Dots was the best at catching a scheduling conflict. That is progress — just not the kind that removes the need for human oversight.