Is Siri an AI agent?

The Question Everyone's Asking: Assistant or Agent?

"AI agent" has become tech's most overused phrase — applied to everything from ChatGPT to a smart thermostat. So when Apple starts calling Siri an agent, skepticism is fair. After all, Siri debuted in 2011, and for most of its life it was the thing you yelled at to set a 10-minute timer.

The confusion is real, and even experts feel it. Instructors at the Swiss Cyber Institute note that "everyone has their own definition of an agent. It varies widely, from the abstract to the extremely specific." Without a shared definition, the label means little.

Here's a plain-language rubric you can apply to any AI tool, including Siri. Systems sit on a spectrum from chatbot to assistant to agent, evaluated on four capabilities:

• Reasoning — can it draw inferences from context, not just match keywords?
• Memory — does it know who you are, your habits, your data?
• Planning
— wait, let's list it properly below.

• Planning — can it break a goal into multi-step workflows across apps?
• Tool use — can it take actions on your behalf, not just answer questions?

The dividing line is simple: an assistant depends on you to drive each step forward. An agent executes the whole task once you state the outcome. As the Swiss Cyber Institute puts it, with an assistant, "the moment you stop prompting, it stops working."

By that standard, old Siri was clearly an assistant — one that, as MindStudio observed, "still couldn't do much more than set timers and play songs." The redesigned Siri is a different proposition. Hands-on testing by tech columnist David Pogue, who ran 125 real-world commands, found compound requests like "find out when the next Mets game is, add it to my calendar, and ask Tracy if she wants to watch it" working reliably — genuine multi-step, cross-app execution.

Apple, for its part, now frames Apple Intelligence not as a chatbot but as a system of integrated, task-specific AI agents, and its 2026 redesign positions Siri to "take actions in apps" based on what you're doing in the moment. With over 1.5 billion iPhones in use, the stakes of that repositioning are enormous.

So what's the verdict? Yes, Siri is now an AI agent — with limits. It plans, it acts across apps, and it reasons about your personal context. But it stumbles outside Apple's ecosystem, leans on ChatGPT for hard problems, and has lost some third-party actions it used to handle. One reviewer's summary captures it well: "a brilliant remote control for your digital life and a mediocre thinker."

If you're extending your own agents' capabilities beyond what any single assistant offers, agent skills are the practical path — SkillzBench is building a free, security-scanned directory of version-pinned skills with a free MCP connector, so you can browse, verify, and equip with confidence.

What Makes the New Siri Actually Agentic

The difference between the old Siri and the new one comes down to a single question: who drives each step? The old Siri "depended on you to drive each step forward," as the assistant-vs-agent framework puts it — you prompted, it responded, then you prompted again. The redesigned Siri flips that dynamic.

At WWDC, Apple shipped three core capabilities that make this possible: personal context, screen awareness, and cross-app orchestration built on App Intents. Together they let Siri understand your situation, see what's on your screen, and act across Messages, Calendar, Reminders, Photos, and more — describing outcomes rather than issuing commands.

The most convincing evidence comes not from Apple's marketing but from tech columnist David Pogue, who ran 125 hands-on tests of the new Siri. The standout results were compound commands like "Find out when the next Mets game is, add it to my calendar, and ask Tracy if she wants to watch it" — which worked reliably, "dazzling every time." That's a single spoken sentence triggering search, calendar action, and messaging in sequence, with no manual handoffs.

Pogue's tests also surfaced contextual inference, arguably the most agentic trait of all. Ask "What time do I need to leave for the wedding?" and Siri deduces which wedding you mean and where you're starting from, then calculates departure time. The system also operates the phone itself: taking selfies, drafting and sending emails, creating automations, and extracting text from images.

• Personal context: Siri reasons over your calendar, messages, and location — on-device, with data "locked on the phone."
• Screen awareness: it acts on whatever you're looking at, in the app you're already using.
• Cross-app orchestration: App Intents exposes typed app actions that Siri chains into multi-step workflows.
• Phone operation: selfies, emails, and automations executed end to end without you opening an app.

The architecture matters as much as the demos. App Intents lets developers expose discrete, callable actions that agents invoke directly, bypassing UI navigation — and Apple has tested expanded intents with Uber, Amazon, WhatsApp, YouTube, and others. Analysts describe the result as an ambient agent runtime rather than a chatbot: Siri as the layer between you and everything on the device.

The gaps are real — third-party actions like calling an Uber regressed, and the Health app isn't integrated yet. But for readers tracking the broader agent ecosystem, the pattern is familiar: capability lives in the skills and intents an agent can actually invoke. That's exactly the problem SkillzBench is building for — a free directory of verified, security-scanned agent skills plus a free MCP connector, so you can equip any agent with capabilities that actually work.

Where Siri Falls Short: The Honest Limits

For every dazzling demo of the new Siri, there's a quieter list of things it simply cannot do — and pretending otherwise does nobody any favors. Independent testing shows the agent is real, but so are the cracks.

The sharpest verdict comes from Fello AI's hands-on comparison: Siri is "a brilliant remote control for your digital life and a mediocre thinker." It executes tasks inside Apple's ecosystem with impressive reliability, but it cannot generate a document or PDF, and it stumbles the instant a request leaves Apple's walled garden. When the reasoning gets hard, it hands the baton to ChatGPT.

David Pogue's 125-test hands-on evaluation confirms the pattern. Compound commands like finding a Mets game, adding it to a calendar, and texting a friend worked "dazzling every time." But roughly ten explicit failures dotted the results, and some of them sting.

The regressions are the surprising part. Because Siri AI requires software companies to adopt new techniques, some commands that worked under the old Siri are now broken. "Call an Uber" no longer works directly, and neither do Venmo payments. The Health app isn't integrated at all, and some in-app actions fall back to showing you instructions rather than doing the task. That's the opposite of what an agent is supposed to do.

Then there's the availability problem:

• Staggered rollout: Apple is releasing Siri AI via a waiting list because, as Pogue notes, it "can't unleash it to 2.5 billion devices all at once."
• EU users get Siri AI only on Mac and Vision Pro at launch — not iPhone, iPad, or Watch.
• Mainland China is excluded entirely for now.
• The beta shipped in English only, with French, Japanese, Korean, Portuguese, and Spanish arriving October 2026 alongside the iOS 27 public release on 14 September.

For business readers, the lesson is broader than Apple. An agent is only as capable as the skills and connectors wired into it — and when integrations regress or go missing, the agent's competence regresses with them. That's why the industry is converging on standards like App Intents and Model Context Protocol, which let any compatible agent act on apps through a shared orchestration layer.

The same logic applies to any agent you deploy. Extending one with vetted, benchmarked skills — the kind SkillzBench is building into its free directory, with version-pinned, security-scanned agent skills and a free MCP connector — is how you avoid the "mediocre thinker" trap. Siri's limits aren't a failure of ambition; they're a reminder that agency is earned integration by integration.

Why App Intents and MCP Change the Game

Strip away the voice interface and Siri's real story is architectural: Apple has quietly built a runtime where software actions become things an AI can call. That infrastructure, not any single demo, is what makes the agent label stick.

The foundation is App Intents, the framework that lets developers expose discrete, typed actions that AI agents can invoke directly, bypassing app UI navigation entirely. Available since iOS 16, the framework expanded significantly when 12 domains launched at WWDC 2024 — spanning Mail, Photos, Books, Browsers, Cameras, Journals, Spreadsheets, Word Processors and more, according to industry analysis. The strategic implication is blunt: apps that don't expose intents effectively don't participate in agentic workflows.

MCP takes this further. Code found in iOS 26.1 developer betas shows Model Context Protocol support built into the App Intents framework — meaning any MCP-compatible agent, whether ChatGPT, Claude, Gemini or a local model, could act on apps through Apple's orchestration layer rather than Siri alone. Reports also indicate native MCP support would let Siri connect to enterprise CRMs, knowledge bases, and ticketing systems.

Apple's positioning is deliberately not a chatbot. As Craig Federighi put it, building a chatbot was "never the goal" — the aim is AI "within reach whenever you need it." Analysts frame Siri as "the layer between you and everything else on the device, an ambient agent you invoke from wherever you already are," running across more than 1.5 billion iPhones in use.

For business readers, the scale matters as much as the architecture. Gartner predicts that by the end of 2026, 40% of enterprise applications will feature task-specific AI agents, per Gartner's forecast. Apple's design decisions preview how that shift reaches consumers:

• Typed, AI-callable actions replace screen-scraping and manual app navigation.
• A neutral orchestration layer lets multiple models act, not just the vendor's own assistant.
• On-device processing plus Private Cloud Compute keeps personal data inaccessible to Apple or anyone else.
• Third-party participation is opt-in — Apple has tested expanded App Intents with Uber, Amazon, WhatsApp, YouTube and others.

The same pattern plays out beyond the phone. Agents everywhere now extend their reach through skills and MCP connectors, and quality varies wildly — published research found roughly 26% of agent skills carry at least one vulnerability. That's why SkillzBench is building a free directory of version-pinned, security-scanned, benchmarked agent skills plus a free MCP tool, so builders can extend any agent with skills they've actually verified. Apple's bet is that orchestration, not any single model, wins the agent era — and the ecosystem forming around MCP suggests the industry agrees.

How to Put This Into Practice

Knowing whether Siri qualifies as an agent matters less than knowing how to evaluate any tool that claims to be one. The four-box rubric from the Swiss Cyber Institute — reasoning, memory, planning, and tool use — gives you a quick, vendor-neutral test. If the system depends on you to drive each step forward, it's an assistant, not an agent, no matter what the marketing says.

Test with compound commands, not single queries. Anyone can answer "what time is it?" — that proves nothing. Pogue's 125-test hands-on evaluation found that compound commands like "Find out when the next Mets game is, add it to my calendar, and ask Tracy if she wants to watch it" worked reliably, while single-step queries reveal nothing about orchestration ability. Run your own version: chain three actions across two apps and see whether the outcome arrives without you prompting each step.

For builders: expose your app's actions via intents. This is the practical takeaway from Apple's architecture. As Nate B. Jones puts it, "Apps that don't expose intents are apps that don't rank" — meaning they simply don't participate in agentic workflows. Pogue's testing showed the flip side: commands like "Call an Uber" stopped working because third-party apps hadn't adopted the new techniques. If your app isn't agent-accessible, agents will route around it.

Extend any agent's real capabilities with vetted skills. No agent — Siri included — does everything. The practical move is to extend agents with skills, but choose them carefully: published research found roughly 26% of skills carry at least one vulnerability. That's why security scanning, version pinning, and benchmarking matter before you trust a skill with real work.

A quick checklist before trusting any "agent":

• Run the four-box rubric: does it reason, remember, plan, and use tools — or just answer questions?
• Test a multi-step, cross-app command and watch whether it completes without step-by-step prompting.
• Check which apps and actions are actually exposed — gaps in integration mean gaps in capability.
• For skills and extensions, verify security scans, version pins, and benchmark results before installing.

SkillzBench is building a free answer to that last step: a directory of security-scanned, version-pinned agent skills plus a free MCP connector, with every skill benchmarked in a sandbox so you can see the Lift it actually delivers. Browse and test free skills at skillzbench.com — because an agent is only as trustworthy as the capabilities you hand it.