What the job ad actually described
According to Technology Magazine's report on the posting, the role sat across three groups at once: Siri, UI Frameworks and Apple Foundation Models. That combination is the interesting part. Siri supplies the intent layer, UI Frameworks supplies the ability to reach into windows and controls, and Apple Foundation Models supplies the model doing the reasoning.
The ad's language was specific about scope. A "general-purpose computer-use" agent, not a task-specific helper. It would coordinate sub-agents for multi-step workflows – the pattern where one model plans and several others execute smaller jobs. The posting also emphasised privacy, sandboxing and permission models, which is what you write when the agent has to be trusted with a file system.
Nothing in the description mentions a ship date. Bloomberg's reporting, cited in the same piece, puts pressure from CEO John Ternus on software chief Craig Federighi to move faster, with a launch window as soon as late 2027.
The posting vanished after two days
Apple pulls requisitions regularly, and a removed listing is not evidence of a cancelled project. It is evidence that someone decided the vacancy was drawing attention it did not want. The removal came after the story travelled from Bloomberg into the trade press.
Markets noticed something, at least briefly. Meta fell nearly 2% and Anthropic dropped more than 3% in after-hours trading following the reports, which says more about how thin the agentic narrative is right now than about Apple's actual progress. Apple has not shipped a shipping agent and has not announced one.
macOS 27.3 raised the cost of allowing a third-party agent
Here is the concrete number worth keeping. In the macOS 27.3 beta, granting a third-party agent – ChatGPT is the obvious example – full computer-use privileges now takes 14 setup steps, up from 4. Apple also added warnings telling developers about the security risks of rogue agents.
Read that as a product decision, not just hardening. A computer-use agent with Full Disk Access can read your mail store, your notes, your keychain-adjacent files and every document in your home folder. Four clicks was thin. Fourteen steps, each with a warning, is a deliberate friction tax on agents Apple does not build.
It also sets up a neat asymmetry. Apple's own agent, built into Siri and the foundation models stack, would arrive with permissions already granted by the operating system. Everyone else gets the setup wizard.
Where the competition stands
Apple is late, and the numbers show by how much. Meta's Muse agent crossed 10 million downloads in 40 days and ranked second among free US apps; ChatGPT sat at third, Claude at fifth and Gemini at seventh. Google's Mac Gemini app already exposes computer-use controls, and OpenAI has Dots in the field.
API pricing for the models behind these agents has settled into three bands, per million input/output tokens: Meta's Muse Ember at $0.10 / $0.50, Mistral Large 4 at $0.70 / $2.10 with free open weights, and the frontier tier – Claude Fable 5.2 at $8 / $40, GPT-6.2 Astra up to $8 / $16 in and $80 out. An agent that runs twenty tool calls per task turns that spread into real money, which is why the cheap tier matters more for agent workloads than for chat.
There is demand to meet. An OutSystems study cited in the same reporting found 93% of IT organisations planning or building custom AI agents. A NowSecure analysis found 24% of examined mobile apps already exposed capabilities to Siri agent workflows ahead of iOS 27.
The European part is not a footnote
Anyone in the EU reading the late-2027 date should treat it as a US date. Deeply integrated OS agent features have been the first thing to slip in Europe: next-generation Siri capabilities have been withheld or delayed here while shipping elsewhere, because the Digital Markets Act's interoperability and data-sharing duties change what a gatekeeper may reserve for its own services.
That cuts both ways for European users. The DMA's self-preferencing ban pushes in the opposite direction from a 14-step permission screen – Apple has to justify why its agent gets access that a rival's does not, and that argument is harder to make in Brussels than in Cupertino. The practical consequence for a Czech or German developer today is unglamorous: build your agent against documented macOS APIs, expect the permission flow to change again before release, and do not design a workflow that depends on Full Disk Access being easy to grant.
Regulatory weight has also shifted. Since 2 August 2026 the European AI Office holds full statutory enforcement powers over general-purpose model providers, including audit rights and fines up to €15 million or 3% of global turnover. An agent that reads files and forwards their contents to a cloud model is a data-protection question for any EU company deploying it, independent of which vendor built the agent.
What to take from a job ad
Apple has not confirmed the product, the date or the price. What is confirmed: a requisition existed, it described a computer-use agent with sub-agent orchestration, and it is gone. What is also confirmed is the direction of travel in the beta – third-party agents on macOS now need fourteen deliberate steps to earn the same access an Apple agent would inherit.
If you run local models and want to know what an agentic loop costs in tokens per second on hardware you own rather than rent, our AI Arena measurements are the place to start. The gap between a demo agent and one you would let touch a terminal is usually measured there, not in a keynote.
Has Apple confirmed it is building a computer-use agent?
No. The evidence is a job requisition posted on 17 November 2026 and removed on 20 November, plus reporting on internal pressure to accelerate. Apple has made no product announcement and given no launch date.
Do the 14 permission steps apply to European Mac users?
The macOS 27.3 beta change is not described as region-specific, so the flow is expected to apply in the EU as well. Whether the DMA forces a different arrangement for rival agents is unresolved.
Will the agent run on-device or in the cloud?
The posting stresses sandboxing and privacy but does not say where inference happens. Apple's existing split between on-device models and Private Cloud Compute is the likely pattern, but that is inference, not confirmation.