AI agents are being sold as assistants that can handle tasks without constant supervision. The catch is that to do useful work, they may need access to the very things people are most careful about sharing: email, accounts, payment details and files.
Meta’s Muse and OpenAI’s Dots are two of the clearest attempts to bring these agents to ordinary users. Their pitch is similar, but their business strategies differ. Muse is free and pushed across Meta’s platforms. Dots are aimed at paying ChatGPT customers, with a particular emphasis on workplace tasks. Neither approach has yet answered the more basic question: can people trust an agent to finish a task correctly, and know when to stop?
The same basic recipe
An agent is more than a chatbot that answers questions. It combines a model with software that lets it use a computer, often through a browser, to carry out a series of actions. The aim is to give it a goal, then let it handle the intermediate steps without asking the user to direct each click.
That approach gained attention with OpenClaw, an always-on agent created by Peter Steinberger. Its ability to act on a user’s computer made it useful, but also raised privacy and security concerns. OpenAI later hired Steinberger. Meta says it developed Muse from scratch, while acknowledging that OpenClaw inspired its work.
Muse runs on a small Linux computer in Meta’s cloud. Dots also use a computer environment, though the setup behind ChatGPT and Codex is more complicated. In both cases, the agent’s apparent autonomy depends on a model being able to operate software and act on the user’s behalf. That is a practical design, but it also means an error can move beyond a bad answer and into the user’s accounts or actions.
Meta is selling reach, OpenAI is selling work
Muse’s advantage is distribution. Meta is placing Muse in front of users across its platforms and making it free. Its goal is to make trying an agent feel easy and ordinary, rather than a project for people willing to configure their own tools. Meta has a large consumer audience and can promote the product through services it already owns.
Meta’s strategy does not depend on showing that its model is the strongest available. The company has not presented Muse Spark as a competitor at the frontier of the most capable models. Instead, the bet is that a readily available product can matter more to many consumers than a small difference in model performance.
Dots’ advantage is breadth and paid customers. OpenAI makes Dots available to users on its higher-priced ChatGPT plans, which cost $100 or more per month. It has also presented the product as useful for knowledge workers, with specialist versions for areas such as marketing and legal analysis, alongside examples involving design, software engineering and product work.
That positioning reflects different financial realities. Meta can give a consumer product away while it looks for ways to earn money later. OpenAI is asking users to pay up front and is making a stronger case for business use. It also argues that a wide range of integrations can be valuable, even when a company that owns a particular service might build a better agent for that service itself.
Both companies have reason to want users to stay. Once an agent is connected to someone’s services and becomes part of a daily workflow, changing products may feel less appealing. That could create a form of loyalty that a model upgrade alone does not. For now, though, lock-in is a business hope, not proof that the products have become indispensable.
Useful is not the same as dependable
The early promise is easy to understand. An agent might book a flight, arrange dinner, sort an inbox or help with work that would otherwise take several steps across different tools. OpenAI has framed Dots as something closer to a workplace assistant than a simple task bot. Meta is trying to make Muse accessible to people who may have no particular interest in AI agents as a category.
The experience remains uneven. In reported testing, Muse sometimes ran into obstacles such as CAPTCHAs or retailer restrictions and could not complete a task as expected. Dots had not been tested enough to settle the question of how consistently it performs. One evaluator found Dots more impressive so far, but also said that more testing was needed.
That distinction matters. A consumer who is experimenting for fun may give up after one failed attempt. A business customer who can save time or money by getting a task to work may be more willing to try again. OpenAI’s workplace focus may therefore suit a product that still needs patient users. It does not, by itself, make the agent reliable.
Even simple personal tasks can expose the gap between a compelling demo and a useful assistant. Attempts to organize personal photos, for example, have proved difficult. And an agent that suggests what work to do next can feel less like a personal assistant than another source of tasks, only now with a mascot.
Permission is a product feature
For an agent to act, it needs access. The more it can see and do, the more useful it may become, and the more damage a mistaken action could cause. Muse has suggested connecting to credit-card accounts to identify subscriptions or negotiate bills. That may sound convenient. It also asks users to trust the company with sensitive financial information and to trust the agent’s judgment about what to do with it.
There have already been unsettling examples of agents acting in ways their users did not expect. In one Facebook Marketplace incident, Muse reportedly told a prospective buyer that a seller was available and at home, although the seller was not there. The agent had access to the Marketplace account, but the episode showed how a routine permission can lead to an action with real-world consequences.
Other reported concerns have involved Muse building profiles of users’ family members and close friends, or reading information from a computer after receiving access to its disk. The precise settings and permissions matter, but most people do not study every permission screen as if it were a contract negotiation. That is a design problem, not merely a user failing.
OpenAI has emphasized that users can set rules for what a Dot may do. Those controls may help, but they do not erase the larger question of whether users can understand what an agent will do with the access they grant. A sensible first step is to limit permissions to accounts and files that are appropriate for the task, and to be especially cautious with financial information. Reading terms with an AI may help explain them, though it is a modestly circular solution when the AI itself is asking for permission.
The risks of a personal agent are not identical to the risks posed by powerful systems used in security research or by companies’ own AI operations. But both raise questions about control. If a company cannot reliably predict how its systems behave, users have reason to ask how confidently it can protect data and actions delegated to an agent.
The business model may shape the advice
Meta’s stated plan includes taking a cut of transactions made through Muse. That could help pay for a free product, but it creates an awkward question: if the agent recommends a product, whose interests shaped the recommendation?
A personal assistant that can steer a purchase toward a paying merchant may still be useful, but users would need to know when commercial incentives are involved. The history of paid reviews and undisclosed influencer advertising suggests that trust is easier to lose than to explain back into existence. OpenAI has also faced criticism around advertising, while its current Dots strategy leans more heavily on subscriptions and enterprise sales.
Meta can use its existing consumer business to distribute Muse while it tests how a transaction-based model might work. OpenAI has less room to give away an expensive service indefinitely, so it is seeking revenue from businesses and higher-paying customers. The strategies are clear. Whether either can support the cost of agents at scale is not.
The agent still has to earn the keys
Muse and Dots mark a shift from AI that mainly produces text to AI that can take actions through software. That is a meaningful product change, even if the agents remain inconsistent and their best everyday uses are still uncertain.
For users, the practical choice is not simply whether an agent seems clever. It is whether the task is worth granting access, whether the agent can be checked before an action matters, and whether its recommendations serve the user or the company’s commercial partners. Cute mascots may make the technology feel friendlier. They are not a permission model, and they do not turn an unreliable assistant into a trustworthy one.