Apple is building a Mac AI agent — its rivals need 14 steps

Apple has spent the past month explaining, in unusually blunt language, that handing an AI agent control of a Mac is dangerous. This week it started hiring the team that will get that control anyway.
Three job listings posted to Apple's careers site on November 17 and 18 describe a group building "general-purpose computer-use agents" that plan and carry out multi-step tasks across apps, browsers, command lines, code and APIs, and coordinate sub-agents along the way. Bloomberg's Mark Gurman reported that Apple has considered shipping an agent as soon as the end of 2027, and that development is early. The senior manager listing reads like a mission statement: "Imagine an AI agent that can use a computer the way a capable person does."
In the same beta cycle, Apple made the same capability harder for everyone else. Enabling ChatGPT's computer-use tools in macOS 27.3 now takes 14 steps, up from four, per Gurman. Those two facts belong in one sentence, because they are one strategy.
Two doors in the same operating system
Apple is opening the assistant layer and closing the machine layer at the same time.
Opening: the iOS 27.3, macOS 27.3 and visionOS 27.3 betas let any developer expose its app's actions to Siri AI through App Intents, with no entitlement and no application process — the change we broke down in Siri AI app intents open to every app in iOS 27.3. Closing: Apple told developers on November 16 that macOS 27.3 beta 2 splits Device Control and Data Access into two permissions — System & App Protection Override and Accessibility — adds a single grant flow that starts by picking the requesting app, pushes users toward specific apps or folders instead of the whole disk, offers an "ask every time" setting, and adds a menu bar indicator when one app is driving another. Granting this access, Apple says, "puts all of the data on their Mac at risk."
So third-party apps get a clean, documented path to act inside Siri. Third-party agents get a warning screen, a folder picker, a repeat-consent toggle and a 14-step course to act on the machine. Both moves are defensible alone. Together they draw a line exactly where Apple's own product will sit.
What 14 steps actually is
It is not a security control. Every one of those steps can still be completed, and Apple's note says plainly that permissions already granted stay granted — every existing ChatGPT, Claude and Gemini install keeps the access it has today. What changed is the cost of the next grant, and who pays it.
The person paying is not OpenAI. It is a user who wants an agent to file receipts and now hits a modal saying the Mac cannot protect their data from the app they are about to authorize. Some will read it and stop. Some will click through fourteen times out of habit, which is worse: consent screens that long stop being consent and become a ritual. The split into per-app and per-folder grants is the genuinely good idea here — an agent that only needs Mail should not read Messages — and the part no competitor can copy, because only the OS vendor can enforce it.
The part that isn't good is the asymmetry. Apple's listing asks its own hire to make "safety, privacy, and security central to the design" for agents that "run code, use credentials, and act on a user's behalf," and to build "strong permission models, sandboxing, and transparency." Fair. But a first-party agent inside macOS does not need to walk anyone through fourteen screens; it can hold an entitlement, the mechanism Apple just handed to backup and assistive-technology developers. Apple has not said it will. It has also not said it won't, and the architecture makes that the path of least resistance.
Who wins and who loses
Apple wins on distribution. Meta's Muse hit 10 million US downloads in 40 days and sits at No. 2 among free apps, according to Sensor Tower figures reported by 9to5Mac, and ChatGPT and Claude are both top-five. Apple does not need to win the agent market on quality; it needs the agent market to arrive on hardware it controls, where the default wins by gravity. That is the Siri playbook, run again on a longer fuse.
Rivals lose onboarding conversion, not access. Nothing here bans a third-party agent; it taxes the installation, and the tax lands hardest on companies depending on casual users trying the feature once. Gemini's computer use on the Mac, which arrived Thursday limited to paid plans, now walks into the same gate.
Developers of backup and assistive tools get a documented escape hatch through entitlements. Agent startups do not, because no entitlement exists yet for "I am a general-purpose agent and I need the machine." Apple says entitlement details are "coming soon" and does not mention agents. That silence is the most important thing in the announcement.
What the skeptics get right
The charitable reading is that Apple is correct about the risk and simply moving first on the hard part. A process with full disk access can read, change, delete and export everything on the machine, and only the OS can bound it. Build the consent model before agents ship at scale and users are better off than with a toggle and a shrug. Apple has also been here before: late, mocked, then default.
The uncharitable reading is that friction is a moat you can describe as a value. Gurman reported that some AI companies and users see the extra steps as an attempt to slow competitors; Apple's answer is security. Both readings survive the evidence, which is why the test is not the wording but the next release: whether Apple's own agent meets the same fourteen screens.
Two cautions. Job listings are not a product, and the senior manager role appears to have been pulled — Digital Trends noticed it missing a week after it was posted. The end-of-2027 date is Bloomberg's reporting, not Apple's plan, and it collides with Gurman's earlier reporting that Siri could act "more like an agent" as early as next year. Those are probably two products: an assistant that follows through on goals, sooner, and a general computer-use agent, later. The listing also names the hard constraint — models on devices "under strict latency, memory, and power constraints," with preferred experience in quantization and distillation. Multi-step computer use is the workload that punishes small models first.
What to watch
The 27.3 public release, which Gurman expects as early as December, will show whether 14 steps survives to the shipping build. Third-party app intents are expected to reach users in the spring, around iOS 27.5 or 27.6 — if Apple ships a friendly path for agents to act inside Siri while the disk layer stays hostile, the moat is visible in one release cycle. The promised entitlement details will settle whether an agent can be first-class at all. WWDC in June is where a first-party Mac agent would have to appear in a keynote to make end-2027 credible.
One more room in the house: if Apple's agent gets machine control that rivals must clear fourteen steps to request, that is self-preferencing, and Europe has a template for what happens next. The Commission has not opened that case. Apple's hiring does not depend on it.
The listings are a hiring plan, not a product. The honest summary is that Apple has decided the agent era is coming and intends to be the landlord rather than a tenant. The 14 steps are not an accident of that decision. They are what it looks like from the outside.
If a first-party agent skips the consent screens its rivals must clear, is that security or self-preference? Write to us via our contact page.







