Apple Is Building Its Own AI Agent That Can Run Your Mac — While Tightening the Screws on Rival Agents
Apple is building its own artificial intelligence agent that can operate a computer on a user’s behalf, according to Bloomberg, which first reported the plan on Thursday 19 November after spotting job listings on the company’s careers website. The agent would browse the web, work across apps, run commands, write and debug code, and choose the right tool for each step of a task, all without being asked step by step. According to Bloomberg, Apple has eyed introducing the service as soon as the end of next year, although the work remains in its early stages and Apple has not confirmed the project publicly.
The report lands at a revealing moment. At the same time as Apple is hiring to build its own agent, the company is making it significantly harder for rival agents to gain the same kind of control over its computers. The next version of its desktop operating system, macOS 27.3, will warn users against granting third-party agents full access to their Macs, and, according to Bloomberg, turning on ChatGPT’s ability to control a Mac in the new beta now takes 14 distinct steps, up from four. Apple is, in other words, building the door and deciding who else gets a key.
For a company that spent much of the past two years being described as behind in artificial intelligence, the move is the clearest sign yet of how the new chief executive, John Ternus, intends to close the gap. Ternus, who took over from Tim Cook on 1 September 2026, has reportedly been pressing software chief Craig Federighi to move faster on AI, and the agent project is the most ambitious piece of that push to surface so far.
What the job listings reveal
The clearest evidence of Apple’s plans comes from the company’s own hiring pages. Bloomberg spotted a listing for a senior engineering manager to lead a proactive, agentic AI assistant, and the wording of the posting reads less like a job advertisement than a product description. “Imagine an AI agent that can use a computer the way a capable person does,” the listing said, according to TechCrunch. “It browses the web, works across apps, runs commands, writes and debugs code, and calls services, and it chooses the right tool for each step. It does all this privately and with the polish people expect from Apple.”
The listing went on to describe agents that “plan and carry out multi-step tasks across apps, browsers, command lines, code, and APIs, and they can coordinate specialized sub-agents along the way,” according to MacRumors. Safety, privacy and security, the listing said, need to be “central to the design” of the technology. The manager would work across the Siri, user-interface frameworks and Apple Foundation Models teams, joining a group that, in the listing’s own words, is already working on “general-purpose computer-use agents.”
The posting has since been pulled from Apple’s careers site, according to Digital Trends, which reported that the page now says the job cannot be found. A screenshot shared online showed a posting date of 17 November, two days before Bloomberg’s report. TechCrunch said it had contacted Apple for more information and had not received a response.
What matters is not only what the listing says but what it implies about Apple’s ambitions. This is not a plan to add a few more tricks to Siri. A computer-use agent is a different kind of product: software that is given a goal and then works out the steps itself, moving between a browser, a document, a terminal window and an editor, deciding for itself which tool fits each part of the job. It is the technology that Anthropic, OpenAI, Meta and Google have spent the past year racing to ship, and it is the technology that has made the personal computer, rather than the chatbot window, the new front line of the AI race.
The other half of the story: macOS gets harder for rivals
Apple’s agent plans cannot be understood without the second half of Bloomberg’s report. While the company builds its own agent, it is tightening the rules for everyone else’s. The macOS 27.3 beta, released on 16 November, adds new warnings before a user hands an AI agent broad control of their Mac, and places folder-level limits on what such apps can reach.
The change follows a notice Apple issued to developers in October, in which the company said it planned to “introduce additional controls to ensure that users who genuinely wish to grant an app this extraordinary level of access can only do so with very explicit user action,” according to TechCrunch. In practice, according to Bloomberg, enabling ChatGPT’s ability to control a Mac in the new beta now requires 14 distinct steps, compared with four before.
Apple’s defenders will say this is simply security. An agent with full control of a computer can read files, send messages, move money and delete data, and a compromised or careless one is a genuine danger. The company has spent a decade marketing privacy as a feature, and warning users before they hand the keys to their machine to a third party is consistent with that pitch. The job listing itself stresses that the agent should do its work “privately”, and Apple has long argued that on-device processing, backed by its own silicon, is the safest way to run AI.
But the timing is impossible to ignore. The restrictions arrive just as Apple is building its own agent, and just as rivals are trying to establish themselves on the Mac. Google only this week extended its Gemini computer-use feature to the Mac, a development this site reported on 20 November, and OpenAI’s ChatGPT agent and Anthropic’s Claude tools are fighting for the same territory. If third-party agents must climb 14 steps of warnings while Apple’s own agent arrives pre-installed, deeply integrated and blessed by the operating system, the contest is not being run on level ground.
That tension is the story within the story. Apple controls the hardware, the operating system, the App Store and the default assistant on more than two billion active devices. When it decides that a new kind of software is both dangerous in others’ hands and desirable in its own, it can shape the market by friction alone, without ever banning a competitor outright. Regulators in Brussels and Washington, who have spent years examining exactly this kind of platform power, will be watching how the macOS restrictions are applied in practice.
Why now: the Ternus era
The agent project is also the first big AI bet to emerge under Apple’s new leadership. John Ternus, the former hardware engineering chief who led the development of Apple Silicon, became chief executive on 1 September 2026, after the succession was announced on 20 April. Tim Cook, who had led the company since 2011, moved to the newly created role of executive chairman. Ternus is 51, has spent more than two decades at Apple, and is described by those who have worked with him as an engineer’s engineer, more interested in how things work than in how they are presented.
According to Bloomberg’s Mark Gurman, Ternus told Federighi in early November that Apple needed to “accelerate” its software roadmap and “move quickly beyond merely making necessary improvements to Siri”. Gurman reported on 8 November that Apple could make Siri function “more like an agent” as early as next year. The new job listings suggest the ambition goes further than a smarter Siri: the agent described in the postings is a general-purpose system that would work across the whole computer, not only through voice commands.
The pressure on Ternus is real. Apple is, by most measures, the one big technology company without a frontier AI model of its own. Its rebuilt Siri, which shipped with iOS 27 in September 2026, is powered, according to reports published in January, by a custom version of Google’s Gemini model, in a deal said to be worth about $1 billion a year, with ChatGPT handling overflow queries through an existing integration. Apple has chosen, deliberately, to stay out of the race to build the biggest models, betting instead that its advantage lies in devices, chips and integration. The agent project is the clearest test yet of whether that bet can work: an agent that lives on the device, knows the user’s files, messages and calendar, and acts across apps is a product only a platform owner can build well, and Apple owns the platform.
The agent race Apple is joining late
Apple is not entering an empty field. In the past few months, every major AI company has shipped a proactive agent, a system that does not wait to be asked but offers to do work on its own.
Meta launched its Muse agent line in September 2026, and the app quickly shot up the App Store charts, according to MacRumors. OpenAI followed with Dots, its own agent for carrying out tasks across apps and the web. Anthropic launched Claude Duet. Google turned its Gemini Spark agent into “Tasks” and added new features, and this week brought Gemini’s computer-use capability to the Mac. Each of these systems promises the same thing: give it a goal, and it will work out the steps, use the tools, and come back with the result.
What Apple brings to this race, late as it is, is distribution. Siri is already on every iPhone, iPad and Mac, and the rebuilt Siri AI that shipped in September already has personal context, drawing on emails, messages, files, photos and notes, and App Actions that let it operate inside Apple and third-party apps. An agent built on that foundation would not need to ask users to install anything, grant permissions or learn a new interface. It would simply appear where Siri already is, and do more.
The market noticed. According to Gizmodo, shares of Meta fell around 2 percent and Anthropic around 3 percent in late trading after Bloomberg published its report on Thursday. Investors read the news the way Apple surely intended: the company with the largest installed base of premium devices in the world is no longer content to rent its AI from others.
Yet lateness carries its own risks. The agent market is moving fast, and the companies already in it are learning in public, collecting data on what users actually ask agents to do and where they fail. Apple, which prizes polish and tends to ship late rather than wrong, will arrive after its rivals have had a year or more to learn those lessons. The history of Siri, which launched in 2011 and then stalled for a decade while competitors pulled ahead, is a warning that being late and being polished are not the same thing.
Apple’s long road to this point
To understand why the agent project matters, it helps to remember how Apple got here. The company announced Apple Intelligence in 2024 with a suite of writing tools, notification summaries and image features, but the centrepiece, a rebuilt Siri that understood personal context and could act across apps, slipped repeatedly. The delay led to a leadership shake-up in 2025, with Federighi taking ownership of Siri and Apple’s AI models, Mike Rockwell taking charge of Siri itself in March 2025, and Amar Subramanya, a former Microsoft and Google executive, joining as vice president of AI in December 2025.
The rebuilt Siri finally shipped with iOS 27 in September 2026, the same month Ternus took over. It requires an iPhone 15 Pro or later, and it is still not available on iPhones and iPads in the European Union or in mainland China, according to Memeburn. In China, where foreign models are not permitted, Apple Intelligence runs on Alibaba’s Qwen model. Each of those caveats is a reminder of how complicated it has become to ship AI at Apple’s scale, across jurisdictions with different rules about data, models and competition.
The agent project inherits all of that complexity and adds a new layer. An agent that can run commands, move files and spend money on a user’s behalf needs permissions that no chatbot has ever needed, and it needs them on a device that holds a person’s entire digital life. Getting the safety model right matters more than getting the demo right, and Apple’s insistence that safety, privacy and security sit at the centre of the design is, on this reading, not marketing language but a statement of the real engineering problem.
There is also a commercial question that Apple has not answered. The company has not said whether its agent will be a feature of Siri, a standalone app, a part of macOS and iOS, or something else entirely. Bloomberg’s report says only that the company has eyed a launch as soon as the end of 2027, and that the work is at an early stage. Between now and then, Apple will have to decide how much the agent can do without asking, how it explains its actions, and what happens when it gets something wrong. Those are product decisions, but they are also legal ones, and the answers will shape whether users trust an agent with their computer at all.
Analysis: why it matters
The significance of Apple’s agent project lies less in any single feature than in what it reveals about where the AI race is heading. For three years, the contest was about models: who had the biggest, the smartest, the cheapest. Now it is about agency: whose software actually does things in the world, on the user’s behalf, across the messy reality of apps, files, browsers and services. A model answers questions. An agent books the flight, files the expense, writes the code and fixes the bug. The company that controls the agent controls the work itself, and that is a far more valuable position than answering questions ever was.
Apple’s entry changes the shape of that contest in three ways. First, it legitimises the agent as the next default interface. When the company that defines what a computer is for decides that computers should be operated by agents, the rest of the industry follows, and users begin to expect it. Second, it raises the stakes on privacy as a selling point. Apple’s pitch, visible in the job listing itself, is that its agent will do this work “privately”, on the device, under the user’s control. If that promise holds, it is a genuine differentiator against agents that run in the cloud and see everything. If it does not hold, if the agent needs the cloud as much as its rivals do, the pitch collapses.
Third, and most consequentially, it puts Apple’s platform power at the centre of the AI debate. The macOS restrictions show how easily a platform owner can tilt the field: not by banning a rival, but by adding friction, warnings and steps. Fourteen steps to enable a competitor’s agent, against zero steps for its own, is a competitive advantage that no amount of engineering can overcome. The European Union’s Digital Markets Act was written for exactly this pattern, and Apple’s conduct in this area will be examined closely, both in Brussels and in Washington, where the company’s App Store and platform rules have already faced years of scrutiny.
There is also a simpler, human question at the heart of the agent race, and it is one Apple is well placed to answer badly or well. People do not hand over control of their computers lightly. An agent that misreads an instruction can delete the wrong files, send the wrong message or spend the wrong money, and the first high-profile failure will do more to shape public attitudes than a hundred polished demos. Apple’s caution, its insistence on polish, its willingness to ship late, may turn out to be an advantage here, if it means the agent arrives trustworthy rather than merely first. But caution has a cost too, and the company that waited too long on Siri knows it.
For users, the practical consequence is straightforward: the Mac is about to become a contested space. In the coming year, owners of Apple computers will be offered agents from Google, OpenAI, Anthropic, Meta and Apple itself, each wanting the same deep access to their machine, each warning, implicitly or explicitly, that the others cannot be trusted. Choosing between them will be less a technical decision than a decision about trust, and trust, once lost, is the hardest feature to rebuild.
What to watch next
Several markers will show whether Apple’s agent project is real, and how fast it is moving. The first is the macOS 27.3 release itself, and whether the warnings and limits on third-party agents survive into the final version unchanged. The second is Apple’s developer conference in June 2027, where the company has traditionally previewed its biggest software bets, and where an agent preview would be the natural announcement. The third is hiring: the listings that Bloomberg spotted are only the visible edge of a team that will need to grow substantially if a late-2027 launch is to be met.
Also worth watching is whether Apple opens the same deep system access to third-party agents that its own agent will enjoy. If the company publishes the interfaces and lets rivals build on them, the macOS restrictions look like genuine security policy. If it keeps the best access for itself, the restrictions will look like something else, and regulators will say so. The same question applies on the iPhone, where Siri’s App Actions already give Apple’s assistant a privileged route into third-party apps.
Finally, watch the rivals. Meta, OpenAI, Anthropic and Google are not standing still, and the agent market is young enough that a single breakthrough, a genuinely reliable agent that people trust with real work, could reorder it overnight. Apple’s entry is a statement that the company intends to compete. Whether it can compete on the rivals’ timetable, rather than its own, is the question that will decide whether this is a turning point or another late start.
Sources
- TechCrunch: Apple may be developing its own AI agent that runs your Mac
- MacRumors: Apple is Building an AI Agent That Can Use Apps and Code
- Digital Trends: Apple wants its own AI agent for the Mac, and a late 2027 launch is already in view
- Gizmodo: Apple Is Reportedly Building Its Own Full-Featured AI Agent