Apple Is Building an AI Agent for Mac While Making It Harder for Outsiders to Get In: 14 Clicks Instead of 4
Apple is tightening the screws on third-party AI assistants while preparing its own. In the macOS 27.3 beta, according to Bloomberg, giving ChatGPT control of a Mac now takes 14 clicks instead of the previous four. At the same time, Apple has been recruiting a leader for an agent that writes code, navigates websites and operates apps; the launch is planned for no later than the end of 2027. Agent developers should start rebuilding their onboarding for narrower permissions now.
AI-processed from Habr AI; edited by Hamidun News
Imagine you sell an assistant app that can control a Mac: open programs, switch between windows, work with files. A user downloads it and now needs to grant access. Previously this took four clicks. According to Bloomberg, in the macOS 27.3 beta it takes 14 for ChatGPT. For any product built on agents, that is a significant hit to conversion, and it also hints at where Apple intends to go next.
What changed in macOS 27.3
The beta shows new warning screens. The system asks the user to think again before allowing a third-party AI assistant to control the computer. In addition, Apple has split full disk access into several separate, stricter permissions, and added its own warning to each one.
There is an important caveat: the new rules apply only to apps that are being granted access for the first time. Apps that are already set up will lose nothing. For now all of this exists only in the closed developer version, and no release date for everyone has been announced. The text of the warnings has not leaked. Leaks say nothing about Accessibility or screen recording, so talk of a hard ban on viewing the screen remains rumor.
Apple's explanation is standard: security. According to TechCrunch, in an October notice to developers the company promised that permission would be given only by someone who consciously wants it, through a "very explicit user action." Some AI companies see this as an attempt to slow down competitors. The backdrop is fitting: the US Department of Justice is already suing Apple over restricting third-party services' access to features and devices.
Apple's own agent
At the same time, a job listing for Senior Engineering Manager, Agentic AI – Proactive appeared online. Apple was looking for a leader for an "agentic AI assistant" that uses a computer the way a person does. It plans and executes multi-step tasks across apps, browsers, the command line, code and APIs, and coordinates specialized subagents. A separate line in the requirements calls for "strong models of permissions, sandboxes and transparency."
According to the description, the project sits at the intersection of three teams: Siri, UI Frameworks and Apple Foundation Models. Candidates are expected to have experience optimizing models directly on the device, meaning quantization and distillation. This suggests the agent is meant to run locally on Apple's chips, not just in the cloud. The job page disappeared from Apple's website shortly after Bloomberg's report.
The project has no name, and no external partner is mentioned. Development is at an early stage, and Apple is considering a launch "as early as possible, by the end of next year," that is, by the end of 2027. The acceleration, according to the report, is driven by new CEO John Ternus, who is pressing Craig Federighi, who oversees software and Siri. The foundation is already being laid: the iOS 27.3 and macOS 27.3 betas are testing the first agentic Skills for Siri AI, which allow control of third-party apps, such as ordering a taxi.
What developers and product managers should do
The market reacted nervously: in after-hours trading following the report, Meta shares fell more than 2%, and investors began reassessing positions in independent agent developers. This is one report and one evening, and it is too early to draw far-reaching conclusions from it.
The practical takeaway is more modest and more useful. Granting permissions is turning into a funnel, and every extra screen costs users. What to do right now:
- Break onboarding into narrow permissions instead of one broad "give us everything" request. Users are more willing to agree to access that is clear and limited.
- Measure how many people drop off at each step of the permission flow, and compare it with the previous four clicks.
- Run your scenarios on the 27.3 beta, especially disk access, since Apple has split it up.
- Explain in plain language why each access is needed, right before the request.
One question remains unanswered: will Apple's own future agent have to go through the same 14 clicks? The job listing also talks about strict permissions, so perhaps the fence is being built equally high for outsiders and for Apple itself. But if its own assistant gets a shorter path, complaints about competition will have a serious argument, and the US Department of Justice will have one more reason to take a closer look.
- Nov 21, 2026Apple's macOS 27.3 beta 2 will let you limit AI agents' access to Mac files
- Nov 21, 2026Apple Is Recruiting 200,000 Volunteers to Test a Health AI Chatbot
- Nov 21, 2026A Developer Replaced the Shuttered Arc Browser With Meridian: 158 Patches on Chromium and MCP
- Nov 21, 2026Sales of AI Wearables in Russia Up 7x, AI Earbuds Up 40x
Key Questions Answered
What changed in macOS 27.3?
The beta shows new warning screens. The system asks the user to think again before allowing a third-party AI assistant to control the computer. In addition, Apple has split full disk access into several separate, stricter permissions, and added its own warning to each one. There is an important caveat: the new rules apply only to apps that are being granted access for the first time. Apps that are already set up will lose nothing.
What are the key details on Apple's own agent?
At the same time, a job listing for Senior Engineering Manager, Agentic AI – Proactive appeared online. Apple was looking for a leader for an "agentic AI assistant" that uses a computer the way a person does. It plans and executes multi-step tasks across apps, browsers, the command line, code and APIs, and coordinates specialized subagents.
What developers and product managers should do?
The market reacted nervously: in after-hours trading following the report, Meta shares fell more than 2%, and investors began reassessing positions in independent agent developers. This is one report and one evening, and it is too early to draw far-reaching conclusions from it. The practical takeaway is more modest and more useful. Granting permissions is turning into a funnel, and every extra screen costs users.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.