An app permission tells you who gets through the door. A personal agent needs rules for what it can do once it gets there, especially when the job spans several apps.

An assistant asked to clear your Friday can do a tidy job on your calendar and still make a mess of your life. Move an internal check-in and you save an hour. Decline a customer meeting, notify everyone on the invite, and book a replacement slot with the wrong person, and you have created work nobody asked for. The assistant may have had legitimate access at every step. That is precisely the problem.

The excerpt from Stratechery’s Apple update makes a narrow judgment: Apple avoids ruining the Mac but leaves the agent problem unsolved, raising the question of whether it can think outside the app paradigm. The excerpt does not specify the permission changes. It does, however, identify the right test for them. A platform can make it clearer which apps can see your data and still leave you unsure what a personal agent can do on your behalf.

Keep the app boundary. Add a boundary around the job. For delegated work, the platform needs to distinguish reading from acting, preparation from commitment, and the user’s instruction from whatever the agent happens to read next. Those distinctions must survive a handoff between apps. Otherwise you can revoke access after a bad call, but you cannot stop the call from being made.

App permissions answer the wrong question at the moment of commitment

Under an app-centered permission model, the decision is framed around an identifiable requester and a protected resource. Should this app be allowed to see my calendar? That is a useful question. It is also a decision a person can make without anticipating every future use of the app. Access has a recognizable purpose, and denying it can be the right answer. The problem comes when the same grant is treated as standing authority to perform any action the resource supports. The ability to inspect a calendar does not settle whether the assistant can decline a meeting. Even a separate permission to edit it would not settle which meetings it may decline, for which task, or whether it can send a note under your name.

Take Friday afternoon. You ask the assistant to find two free hours. It can inspect your schedule, flag a flexible meeting, and propose a move. If it edits an event, the consequences depend on whose event it is and whether the edit will notify anyone. A private hold is different from an invitation with clients on it. Those distinctions matter more to you than whether the underlying operation is labeled a calendar edit. If you have to know how each app implements a change before you can judge a permission prompt, the platform has asked you to do the wrong work.

That is the useful reading of the criticism in the Stratechery excerpt. The app remains a sensible place to decide who can handle the data. It is a poor place to describe the full scope of a delegated job, because the job may begin with one app, involve another, and finish with a commitment to someone outside both. The right question at the point of action is whether the user delegated this move, under these conditions. Put it in terms a person can check: prepare the new times, preserve the customer meeting, and ask before notifying attendees. The permission has now described a result the user can recognize. A button granting calendar access cannot do that, no matter how clearly it is labeled.

The trust boundary runs through the task, not around the app

Apply the Trust Boundary Model to an ordinary request: review the documents for my meeting and draft a reply. The user supplies the instruction. The documents supply material for the work. The draft is an output for the user to review. Sending it would be a separate action on the user’s behalf. The platform should not assign equal authority to everything the assistant sees along the way. A request in a meeting document might be relevant to the reply. It cannot, by itself, authorize the assistant to send anything. The distinction holds even if every app involved is legitimate and the person who wrote the document is someone you know.

There is another boundary at the recipient. The user may want the assistant to consult a private note while drafting a message to an external partner. Permission to read the note and permission to send the message do not, taken together, give the assistant permission to disclose the note. Nor is an action check complete if it sees only that a message is going to an approved contact. What leaves matters. So does the change from a private working draft to a communication that represents you. Put both facts in front of the person reviewing the send. A prompt that says the app wants access to Mail cannot convey either one.

For an assistant working across apps, a useful platform rule is to carry the same task limits through each handoff. Reading a document does not enlarge the job. Opening a different app does not reset the terms. A draft handed off for delivery should still be subject to the user’s rule about who can receive it and whether a send needs approval. That is a design requirement, not a claim about what Apple’s update implements.

This also changes what to count when you assess the attack surface. The list of connected apps matters, but so do the routes from material the assistant reads to decisions it can execute. A narrow connection that can publish a statement may deserve tighter controls than broad read access used to summarize private notes. Read access still carries privacy risk. The point is to inspect disclosure and action separately, then inspect the places they meet. That is where two reasonable grants can combine into an outcome the user never authorized.

Put the approval where the consequences change

A consent prompt is useful only if the person can tell what saying yes will do. Ask too early and the user must approve a decision the assistant has not yet made. Ask at every step and the user must supervise the process they wanted to delegate. Place the check at the point where the consequences change: a draft becomes a sent message, a tentative plan becomes a reservation, or a private edit becomes a notification to other people. Until then, give the assistant enough room to do the preparatory work. Make the approval about the effect rather than the app it happens to open.

Consider a request to arrange dinner for a visiting friend. Searching options and drafting a message are useful without committing either person. A booking may involve a deposit or a cancellation term. A message may reveal who is traveling or when. A good review screen would say what the assistant intends to reserve, under whose name, on what terms, and what it intends to tell the friend. If the restaurant changes the terms after approval, the assistant should not treat the old approval as a blank check. Bind the approval to the terms the user saw. If those terms change in a material way, show the difference and ask again.

Not every consequential step needs a fresh tap. The Capability vs. Controllability Frontier is an argument for choosing where to delegate, not for keeping every agent in drafting mode. A person might allow automatic moves of private calendar holds within working hours while reserving approval for any change that notifies attendees. The rule is useful because it separates cases by effect. It also gives the person a concrete reason to raise or lower autonomy: the assistant can handle a known class of actions without inheriting authority over everything else in the app.

Provide a way to stop the remaining work. If the assistant has staged a batch of actions and the user withdraws permission, the actions that have not happened should stop. Then show what already happened. Undo a private edit where possible; do not call a message reversible because you can delete the local copy. Review, cancellation, and undo solve different problems. A platform that collapses them into one stop button leaves the user guessing which commitments still stand.

Action controls only count if the destination honors them

A polished assistant can promise to ask before sending. That is a useful interface promise. It is not, by itself, an enforceable permission. For the guarantee to mean anything, the place that carries out the send must be able to reject it without the right approval. Otherwise the rule depends on the assistant remembering it and applying it correctly every time. The same principle applies to a booking or a shared calendar edit. Put the final check where the commitment is made. Keep the explanation in the assistant, where the user can see why work stopped and what it takes to resume.

This is where retaining the app boundary pays off. The destination knows details the assistant’s task view may miss. A calendar app can tell whether an edit changes a private hold or sends an update to invitees. A service that takes a reservation can tell whether accepting it incurs a fee. The platform’s role, in this design, is to carry the user’s limits across the job and make them intelligible. The destination’s role is to describe and enforce the effect. Neither should silently assume the other checked. A reassuring confirmation is worse than useless if the action applied differs from the one it described.

The hardest product choice is what to do when the available control is coarse. Imagine a service that lets the assistant operate through an account with wide privileges but offers no reliable distinction between preparing an action and committing it. Either there is a trustworthy way to preserve the user’s limit, or the assistant should stop short of the commitment. It can give the user a prepared handoff instead. That is less convenient than automatic completion. It is also more honest than advertising draft-only behavior while relying on a broad grant that permits a send.

Apple’s app paradigm and an agent-centered permission layer therefore need not be rivals. Think of the latter as a contract spanning the former. The user can still revoke an app’s access entirely. For an active task, the user can narrow what the agent may do without disabling every useful connection. The essential test is whether the limit survives the trip from the place you state it to the place that changes something. If an app or service cannot honor it, the assistant has reached the boundary of its authority, even if it has not reached the boundary of its capabilities.

The winning platform will make delegated authority legible

The strategic question raised by Stratechery’s Apple coverage is whether Apple can think outside the app paradigm. For personal agents, a concrete version of that question is whether the platform can give people one reliable account of a job that spans several apps. Before execution, show what the assistant is allowed to do. During execution, show what decision it is waiting on. After execution, show what it changed, what information left, and what it did not do. Those views should agree. A history of app openings or generic permission grants would not explain why a customer received a cancellation.

This is a reason to take the Harness Hypothesis seriously. In this case, the harness is the part of the product that turns an assistant’s proposed steps into actions in the world. A model that identifies the right meeting to move is valuable. For a user deciding whether to hand off the job again, it may matter more that the system can distinguish a proposed move from a sent cancellation and can prove which happened. That distinction gives the platform a way to compete on the reliability of delegation, even when the user cannot judge which underlying model made the better plan.

Start with a test a consumer can run without reading a security manual. Give an assistant a real task with an explicit limit: find a better time for our Friday meeting, but do not notify anyone. Look for a proposed change and a clear indication that nothing was sent. Then approve the change for the internal attendees only and see whether the assistant distinguishes that approval from permission to notify everyone else. Finally, ask what remains pending and whether you can cancel it. These are different questions. A product that answers only with a list of connected apps has told you what it can reach, not what you can safely delegate.

That is the bar for Apple and for any platform asking people to run a personal agent. Keep the ability to say no to data access. Add the ability to say yes to a job without signing away every decision that job might involve. The boundary to defend is the user’s authority to make a commitment. Design the permissions around that.

/Sources

/Key Takeaways

  1. Keep app permissions for data access. Scope delegated actions to the task, recipient, and expected effect.
  2. Let agents prepare work without repeatedly interrupting the user. Check authority when a plan becomes a commitment.
  3. Require the app or service carrying out the action to honor the limit. If it cannot, hand the decision back to the user.
  4. Show what happened, stop what is still pending, and distinguish what can be undone from what has already left the user’s control.