An Apple AI agent that runs your Mac for you is on its way. According to Bloomberg, Apple is hiring a senior engineering manager to build a proactive assistant that writes code, browses the web, controls apps and navigates your devices on your behalf. In the same week, Apple added new warnings to macOS for anyone about to hand a third-party agent the keys. Siri, take the wheel? Not without a few questions first.
For people juggling several calendars, this is the part worth watching. An agent that can drive your apps can drive your calendar app too. The interesting question is not whether it can, but when it should ask before it does.
What Apple is building
Bloomberg reported on 19 November that Apple is working on an agent that can plan and carry out multi-step tasks across apps, browsers, command lines, code and APIs, coordinating specialised subagents along the way. That puts Apple in the same race as OpenAI, Anthropic and Meta, whose agents can already act on a Mac.
Which is exactly why Apple is also tightening the rules. As 9to5Mac reports, the macOS 27.3 beta splits the old “Device Control and Data Access” setting in two. Assistive technologies get their own Accessibility permission. Everything else that controls other apps goes under “System and App Protection Override”, which now lets you:
- pick which apps an automation tool may control (Mail, say, but not Messages);
- make the tool ask for permission every time, instead of holding access indefinitely;
- limit it to specific folders rather than the whole Mac;
- see a menu bar indicator whenever one app is controlling another.
Apple doesn’t block agents from asking for this access. It just makes the request louder, and explains that the override exposes your personal data. Apple has been asking “Are you sure?” since the first time someone tried to empty the Trash, so at least it’s on familiar ground.
When should an Apple AI agent ask first?
Strip away the settings screens and Apple has given you one real choice per app: standing access, or ask me every time. Both are wrong for some jobs.
Ask every time, and you get consent fatigue. By the fortieth pop-up, people click “Allow” without reading, which is worse than no pop-up at all. Grant standing access to everything, and one confused or hijacked agent can do a lot of damage before anyone notices. The useful line is not between agents. It’s between kinds of action.
How we drew that line building Calendrz
We had to answer the same question building Calendrz, because Calendrz is itself an always-on automation with standing access to your calendars. That’s the point of it. When a meeting lands in your client’s Microsoft 365 calendar at 11pm, auto-blocking marks that time as busy in your Google Calendar straight away. Nobody wants a pop-up at 11pm asking whether a busy block may be created. So availability mirroring runs at any time, without asking.
What makes that acceptable is how narrow the job is. Calendrz mirrors when you’re busy across your connected accounts; the blocks it creates carry a label you choose, not the private details of the meeting behind them. A small, boring, predictable job is a good candidate for standing access.
Agents are a different matter, and our MCP server treats them differently. Instead of letting an assistant click around our screens, we give it a short list of named tools. Most of them are read-only: show my events, list my connected accounts, check my settings. A few change things, such as updating your mirror preferences or running availability mirroring now. And one or two are destructive: removing a connected account deletes all the busy blocks mirrored from it. That tool’s own description carries a warning telling the agent to confirm with you before calling it.
We don’t leave it at asking nicely, though. When an assistant in your browser uses Calendrz through WebMCP, the destructive actions open our own confirmation dialog, which the agent can’t click through for you. The agent also never sees your Google or Microsoft sign-in tokens, and the access it does get expires after an hour.
That’s the same split Apple is reaching for: standing access for the narrow and reversible, ask first for the broad and irreversible.
What to take from it
- Sort by action, not by app. Reading your calendar and deleting a connected account are not the same risk, even if the same agent does both.
- Keep standing access for boring jobs. If a job is narrow, predictable and easy to undo, let it run. If it isn’t, make it ask.
- Prefer agents that use a front door. An agent driving your screen needs broad access to your whole machine. An agent calling a product’s own tools only gets what those tools allow.
That last point is where this is heading. The calendar screen you click through today will matter less every year, because your agent will do the clicking. The products that win will be the ones that give agents a proper front door, with clear tools and clear limits, so nobody has to hand over the keys to the whole house just to book a meeting.
Whatever agent ends up running your week, it can only respect your busy time if every calendar shows it. Connect your calendars to Calendrz and let availability mirroring keep them honest 🙂







