Joe Rossignol, reporting for MacRumors: Siri AI is much more intelligent and personalized compared to classic Siri, but it is still not nearly as advanced as proactive, always-on agents like OpenAI’s Dots and Meta’s Muse. That may change, though, according to a new rumor.
In his Power On newsletter today, Bloomberg’s Mark Gurman said he was told that Apple could make Siri function “more like an agent” as early as next year. However, if this plan moves forward, the exact implementation remains to be seen.
“Dots, Muse and Grok can keep track of goals, determine what needs to happen next and continue pursuing objectives without the user initiating every individual step," said Gurman. “Apple ultimately wants to go in this direction, I’m told, and I believe the company could make Siri more like an agent as early as next year.”
Further reporting from Gurman: Behind the scenes, though, Ternus has delivered a different message, knowing full well that Apple remains far behind rivals. In recent weeks, Ternus held a closed-door meeting with Craig Federighi, who became Apple’s AI czar following last year’s dismissal of its former AI leader. Desiring more and frustrated by what he sees as a reluctance to embrace AI with the same urgency as competitors, Ternus pushed his software chief to accelerate the company’s road map and move quickly beyond merely making necessary improvements to Siri.
I believe we’re already seeing the fruits of this effort in Apple’s latest beta operating systems. The 27.3 betas include an update to Siri AI’s models and architecture that extends its App Intents capabilities beyond a limited subset of app categories. In the public version of Siri AI, the model can only call App Intents that conform to an app schema domain. The full list of available domains is: audio, calendar, camera, clock, mail, maps, messages, notes, phone, photos, reminders, in-app search, video, books, browser, files, journaling, presentations, spreadsheets, whiteboards, and word processors. That’s a near-exhaustive list of most apps on the App Store, but it’s not all apps. For instance, my app Citations doesn’t fit into any of these categories, and so Siri AI couldn’t call my app’s intent until the 27.3 release.
The 27.3 betas expose App Intents as tools to Siri AI, coupled with optional, developer-written skills — just Markdown files included within the app’s bundle — to tell the model how to use those intents. The new models can also call multiple intents sequentially if a user’s task requires it. In total, Siri AI is a much more capable agentic system now, just a few months after its initial release. It’s difficult to describe what a departure this is from the Apple of just a few years ago, which seldom added big features between major OS versions. The Siri AI models seem to be getting regular updates and are adding new, interesting capabilities, even if they’re considered to be in “preview.” In my review of Siri AI over the summer, I remarked that such ambition would be requisite for any successful artificial intelligence product, and I’m happy to see it.
Back to the report. I do see agentic tools like Muse and Dots as an overzealous frenzy, but I think Apple will be able to overcome this due to its platform advantage. Apple has been proactively recommending apps, photos, and other content throughout iOS for over a decade now, and it’s easy to envision those Siri Suggestions, as they’re called, constructed into a more formal product. They’re only possible because of how much user data Apple already has; other companies must ask for it. And it seems more than plausible that with this newfound focus and ambition — coupled with software engineers armed with agentic coding tools — Apple can ship such features by next year. Relatedly, when the new Siri was first announced in 2024, I thought App Intents gave Apple a much-needed head start in agentic AI. Instead of relying on computer-use models or virtual machines, like Muse and Dots do, App Intents will unquestionably be the marquee advantage of this agentic Siri product, and that ought to make it much more compelling and less gimmicky than Muse or Dots.
Gurman has argued that an Apple-made computer-use model is necessary, but I disagree. I’ve always regarded computer use as a crude stopgap solution between true interfaces made for agents. Software we humans use and what agents interface with should be discrete, and that trend has transpired over the past year in the form of command-line interfaces, coupled with broader support for the Model Context Protocol, as I predicted. Apple’s version of MCP is the App Intents framework: it offers support for user-facing snippets and broad agent-friendly backend integration. I do think that Siri AI ought to be able to interface with the local command line in future versions of macOS, though, similar to Codex or Claude Code. Perhaps the models are not proficient enough at coding, but they should be able to reorganize files, fetch data from the web, and call other command-line utilities. This, coupled with App Intents, is a much more elegant solution to computer use that other tech companies are practically begging for.