The Race to Make Personal AI Work

I tried using the Instacart app in ChatGPT today. It didn’t really work. After spending more time on it than I wanted to, I gave up and did the shopping myself.

I wanted it to work. Ordering groceries is exactly the kind of task I’d like to hand over. But when getting help takes more effort than doing the task myself, I’m going to do it myself.

Meanwhile, Muse reached number one on the U.S. iPhone App Store’s free-apps chart. People are interested enough to try a personal agent. Getting them to keep using it is a harder problem.

That’s the race Meta, xAI, and OpenAI are running with Muse, Grok Bot, and dots. They’re trying to make handing over a task feel worth it.

You still have to explain the task, give the agent access to your accounts, and check its work. That can be a fair trade. But if you spend the whole time answering questions and correcting mistakes, you’re still doing much of the job.

An agent starts to become useful when you can give it a job and spend less time getting to a result you’re happy with. After enough good experiences, you might feel comfortable giving it more responsibility.

We’re still figuring out how to get there.

Even a grocery order involves judgment

Getting the order placed is only part of grocery shopping. If the substitutions are wrong or something doesn’t fit your diet, you’re left with groceries you didn’t want. Having the agent ask you about every item would get tedious too.

Think about the decisions you make while filling a cart. You might pay extra for one brand and buy whatever’s cheapest for something else. Even choosing bananas depends on when you plan to eat them. You already know these things about yourself; explaining all of them to an agent could take longer than shopping.

Now extend that to travel, household spending, or a message to someone you care about. The agent needs to understand what matters to you and recognize when a decision is yours to make. You also need confidence in how it uses the private information that helps it develop that understanding.

In January 2026, Gartner surveyed 322 U.S. consumers and found that 31% would let AI narrow their choices for household supplies.

Consumers were even less willing to let AI make purchase decisions. That figure reached only 11%, even in lower-stakes categories.

The extra checking showed up in Gartner’s November–December 2025 survey too. Of the respondents who’d used AI while shopping for a recent purchase, 54% said they had to check all the information it gave them. And 62% said that information ended up wasting their time.

NMI’s survey of 1,000 U.S. adults found that 70% wanted to be able to review or override an AI decision before a purchase. Seventy-six percent were uncomfortable with AI completing purchases on their behalf.

These surveys asked people about their experiences and what they’d be comfortable doing. They don’t tell us how well a particular agent works today. What I recognize in the responses is the wish to get some help while keeping the final say over what gets bought.

A useful agent could do the research, prepare the purchase, and show me the few choices that need my attention. That would already be worth something.

The delight comes when it gets something right without making you explain every detail. The confidence comes from seeing it do that consistently, including knowing when to ask.

Meta is working on remembering and following through

Meta’s October 12 announcement of Muse Wonder focuses on two abilities that matter here.

It calls the first Active Memory: retaining context across conversations and bringing back the details relevant to the current task. The second is Active Focus: continuing through a difficult assignment, trying another approach when something fails, and checking in before important actions.

Take a food restriction. Once you’ve told the agent about it, that information needs to affect the next order too. An out-of-stock item shouldn’t leave you doing the search again. The agent could suggest a replacement that fits the restriction and ask whether you want it.

Meta’s demonstrations include planning a run around someone’s circumstances, arranging a wedding seating chart, and building a playable 3D game from images. These are company demonstrations, so they don’t tell us how reliably the models will handle everyone’s requests. They do show the range Meta is pursuing, from personal planning to software creation.

Wonder is available through Meta’s Model API and Muse Code, with Wonder Full available to Muse Code subscribers. Meta says Wonder will also power experiences in the consumer Muse agent, starting with subscribers. That rollout is still described as forthcoming.

Read next: VS Code Agents Can Now Work While You're Away

Recommended

Recommended

VS Code Agents Can Now Work While You're Away

You can now ask a VS Code agent to check your repository every morning without starting the session yourself. On my machine, the Automations page in VS Code 1.140 offered a template called Catch up on main, set to run ev…

Read next

Coding is part of the same race

Meta offers Muse Code, SpaceXAI has Grok Build and Cursor, and OpenAI has Codex as well as dots. Their plans for agents extend into building and fixing software.

They’re asking people to delegate increasingly substantial work: build a feature, investigate a bug, prepare an analysis, or handle a task across several apps. The interfaces and permissions differ, but the ambition extends across both creating software and using it.

A proposed bug fix comes with code you can inspect. If you can reproduce the bug, you can try the same steps after the change and see whether it still breaks. That won’t expose every bug or security issue, but it gives you a way to challenge the agent’s claim that the job is done.

Personal decisions can be harder to evaluate. A passing test can help establish that software behaves as intended. It can’t establish that an agent chose a vacation you’d enjoy or understood a preference you barely knew how to explain.

And trust doesn’t automatically transfer between those jobs. I might be comfortable letting an agent prepare a code change I can review while still wanting to approve every purchase.

That’s why having more capabilities only gets these companies part of the way there. They still have to earn permission to use them in each part of someone’s life.

More access can mean more checking

OpenAI describes dots working across projects using their own cloud computers and connected apps. Grok Bot also works across apps and websites and can continue an assignment while your laptop is closed. Both provide controls over access and approvals.

If every connection creates another stream of decisions for you to supervise, the product can become exhausting even as it becomes more capable.

An agent doesn’t need to understand everything about us before it can be useful. It needs enough context to do the assignment, clear limits on what it can decide, and a sensible way to handle what it doesn’t know.

These companies have reasons to keep trying

Meta can put Muse in front of people through WhatsApp. OpenAI can introduce dots to ChatGPT users and make them reachable through Slack and Teams. SpaceXAI brings Grok, Grok Bot, Cursor, and X together in one subscription.

That gives them repeated opportunities to get people to try again. They can also improve several parts of the experience together: the model, the computer it runs on, the connections to other services, and the app where you ask for help.

There’s a lot to gain if that works. An agent you regularly consult could influence which products you consider, which services you use, and what information you see before making a decision.

It also means the company asking for more access to your life may become the company you depend on to make sense of it. I’d want to know what it remembers, what it did on my behalf, and what happens if I decide to leave. If something goes wrong, I’d want a way to undo it and a clear answer about who can help.

Muse’s App Store ranking shows that personal agents are getting attention. My grocery order shows how quickly a willing user can decide to take the work back.

I still want this to work. I’d like to hand over the shopping and trust that the result will be close enough to what I wanted, with a few decisions left for me where they matter.

For now, doing it myself was easier. These companies are racing to change that.

---

References