WWDC is Apple's developer conference, and in recent years it has become one of the most-watched moments in consumer AI, because Apple's choices shape what billions of phones can do by default. The interesting question for anyone working on AI agents is not which on-stage demo looked best. It is structural: what does Apple's direction do to the layer where agents actually live, the layer where software takes real actions across real apps on your behalf.

This piece is deliberately grounded. Rather than narrate specific announcements, it reads the trajectory Apple has been on since it introduced Apple Intelligence at WWDC 2024, and reasons forward about what an on-device, app-intent world means for autonomous agents. The verifiable building blocks, on-device foundation models, the App Intents framework, Private Cloud Compute, and a more capable Siri, are already public. The implications for agents follow from how those blocks fit together, and that reasoning is the durable value here.

The Apple Intelligence baseline

Apple introduced Apple Intelligence at WWDC 2024 as a personal intelligence system built around three ideas: on-device foundation models that run locally on the device, a privacy-preserving server tier called Private Cloud Compute for heavier requests, and deeper integration between the assistant and the apps you already use. Apple has documented these pieces for developers, including a Foundation Models framework that lets apps tap the on-device model and the App Intents framework that exposes app actions to the system (developer.apple.com, retrieved 2026; App Intents documentation, retrieved 2026).

That baseline matters because it tells you what kind of AI Apple is building. It is not a single giant chatbot. It is an intelligence layer woven into the operating system, designed to run as much as possible on the device and to act through structured app actions rather than free-form text alone. Apple has been moving steadily in this direction, and the likely trajectory at WWDC 2026 is more of the same: a broader on-device model, more app actions exposed, and a more capable assistant. The agent-relevant question is what that architecture enables and what it does not, which is where the rest of this piece focuses.

On-device models change the cost curve

The single most consequential idea in Apple's approach is that a capable model runs locally, on the phone, at no marginal cost and with no network round trip. That changes the economics of an entire class of tasks. Summarizing a long message thread, ranking which notifications matter, drafting a one-line reply, extracting a date from an email: these are small, well-bounded jobs that a compact on-device model handles quickly and privately. When that work is essentially free and instant, it gets used everywhere, because the friction disappears.

This is not the same as on-device models replacing cloud agents. A small model that fits on a phone is excellent at the light tasks above and weak at the heavy ones: long multi-step reasoning, planning across many tools, and orchestration that holds a lot of state. Those still favor larger models running in the cloud. The honest way to read 2026 is as a tiered split, not a takeover. On-device handles the light, private, high-frequency work; the cloud handles the heavy reasoning. Understanding why that division exists starts with what an agent actually is, which we cover in what is an AI agent.

App Intents is the agent surface

If you care about agents, App Intents is the part of Apple's stack to watch. An agent's value comes from doing things, not just describing them, and to do things it needs a way to invoke real actions in real apps with the right parameters. App Intents is exactly that mechanism: a framework where an app declares its actions and content in a structured form the system can understand and call (App Intents documentation, retrieved 2026). When an app exposes a clean set of intents, the system assistant can trigger them on the user's behalf.

Read at the right altitude, App Intents is a per-app tool catalog. Each declared intent is a tool with a name, inputs, and an effect, which is conceptually the same shape as the tool definitions an agent uses to act. That framing connects Apple's world to the broader agent ecosystem, because cross-platform protocols express the same idea in a vendor-neutral way. If you want the contrast between an agent and the action-exposing layer it calls, our explainer on AI agent vs MCP server walks through the distinction: the protocol exposes tools, the agent decides which to use and when. App Intents is Apple's version of the tool-exposing side, scoped to its own apps.

The strategic point is that whoever controls the action surface shapes what agents can do on a platform. Apple making app actions easy to declare and easy to invoke is good for any agent operating inside the Apple ecosystem, because it grows the catalog of things that can be done programmatically. It is also a reminder that an agent is only as capable as the actions it can reach.

Siri as a bounded assistant

Siri sits on top of all this as the user-facing assistant, and Apple has been working to make it more action-oriented by leaning on App Intents so it can do more than answer questions. The likely direction is an assistant that can chain app actions on your behalf inside Apple's ecosystem: find a message, pull a detail, set a reminder, open the right screen. That is meaningfully closer to agent-like behavior than the command-and-response Siri of a few years ago.

But a device assistant is bounded by the platform it lives on. Siri is strong at orchestrating Apple and well-integrated third-party apps on your iPhone and Mac. It is not designed to log into a web SaaS tool you use at work, run a multi-step process across three different services, and hand back a finished deliverable. That cross-platform, cross-app work, spanning the web and business tools that have nothing to do with the device, is a different job. The boundary is not a flaw; it is a scope decision. It just means a powerful device assistant and a cross-app agent platform are solving different problems.

The privacy versus capability trade-off

Apple's privacy posture is the other piece that ripples outward. The design principle is to do as much as possible on the device and to route anything heavier to Private Cloud Compute, a server tier Apple describes as processing data only for the request and not retaining it. Whatever the exact details at any given WWDC, the direction is consistent: keep data close to the user and make the cloud tier auditable.

For agents, this sets a new floor on expectations. Once a flagship platform normalizes the idea that an assistant can read your data without storing or leaking it, users start to expect the same from every agent that touches their information. That is healthy pressure. The trade-off is capability: the most powerful general-purpose reasoning still happens in large cloud models, so there is an unavoidable tension between maximal privacy and maximal capability. The mature answer is not to pretend the tension away but to be explicit about what runs on-device, what runs in a controlled cloud, and why, so a user can make an informed choice. Agent platforms that handle sensitive work will be judged on exactly that clarity.

Two layers, not one winner

The tidy way to summarize Apple's direction is that it strengthens the device-and-ecosystem layer of AI: cheaper light tasks on-device, a richer catalog of invokable app actions, a more capable assistant, and a higher privacy baseline. That is a real and useful layer. It is also not the whole stack.

Above it sits a cross-app layer: the orchestration of work that spans web services, business software, and tools that have no presence on your phone at all. A user does not stop needing a research report compiled from five web sources, a spreadsheet reconciled against a database, or a workflow run end to end across SaaS tools just because their phone got better at summarizing texts. These two layers are complementary. A stronger device layer makes the in-ecosystem experience better; a strong cross-app layer covers everything the device assistant was never scoped to touch. This is the same dynamic we traced in our read of what Google I/O 2026 means for AI agents: the platform owners reinforce their own ecosystems, while general cross-app orchestration stays an open field. It also feeds the bigger question of whether agents eventually subsume traditional software, which we take up in will AI agents replace SaaS tools.

What it means for buyers and builders

If you are choosing or building agents, a few practical conclusions follow. First, do not read a strong Apple keynote as the end of the cross-app agent opportunity; read it as the device layer maturing. The work that lives off the device, across the web and business tools, is largely untouched by a better Siri. Second, expect the App Intents pattern, structured, declarative app actions, to keep spreading, which is good news because it makes more of the world programmatically actionable for whatever agent is calling it. Third, treat privacy as a design requirement, not a feature, because the baseline just moved.

This layered view is also why Gravity is built the way it is. Gravity runs expert-built agents that complete an outcome end to end across web apps, SaaS tools, and services, the cross-app layer a device assistant does not reach. You describe what you need in plain words and a maintained agent does the work, paying per use. A stronger on-device layer from Apple and a strong cross-app platform are not rivals; they sit at different altitudes and add up. For a wider read on who is doing what across the field this year, see our mid-year AI agent platform rankings for 2026.

Frequently asked questions

What is Apple Intelligence and how does it relate to AI agents?

Apple Intelligence is Apple's personal intelligence system, introduced at WWDC 2024, that combines on-device foundation models with a privacy-focused server tier called Private Cloud Compute. It relates to AI agents through App Intents, the framework that lets apps expose their actions to the system so that Siri and intelligent features can invoke them. That action surface is the connective tissue an agent needs to actually do things across apps, rather than only generate text.

What are App Intents and why do they matter for agents?

App Intents is Apple's framework for declaring an app's actions and content in a structured way the system can understand and invoke. It matters for agents because an agent's value depends on being able to take real actions in real apps. When an app exposes a clean set of intents, a system assistant can call them with the right parameters. App Intents is, in effect, a per-app tool catalog, which is the same idea that protocols like MCP express in a cross-platform way.

Do on-device models replace cloud AI agents?

No. On-device models and cloud agents solve different parts of the problem. On-device models are excellent for fast, private, low-cost tasks that fit a small model: summarizing a message, ranking notifications, drafting a short reply. Larger reasoning, long multi-step workflows, and cross-tool orchestration still benefit from more capable cloud models. The realistic 2026 picture is a tiered split, with on-device handling the light and private work and the cloud handling the heavy reasoning.

How does Apple's privacy approach affect AI agents?

Apple's approach keeps as much processing as possible on the device and routes anything heavier to Private Cloud Compute, which is designed so that data is used only for the request and is not retained. For agents, this raises the bar on privacy expectations: users increasingly expect that an agent reading their data will not leak or store it. The trade-off is capability, since the most powerful general models still run in the cloud, so agent platforms have to be explicit about what runs where and why.

Will Siri become a real AI agent?

Apple has been moving Siri toward deeper, action-oriented integration with apps through App Intents, which is the foundation an assistant needs to behave more like an agent. The likely direction is an assistant that can chain app actions on your behalf within Apple's ecosystem. Even so, a device assistant is bounded by the platform it lives on, so cross-platform work that spans web services, business tools, and non-Apple apps remains the domain of dedicated agent platforms.

What does Apple's direction mean for cross-app agent platforms like Gravity?

It is mostly complementary. Apple is strengthening the on-device, in-ecosystem layer, where actions inside iPhone and Mac apps become easy to invoke. Cross-app platforms operate at a different layer, orchestrating work across web apps, SaaS tools, and services that a device assistant does not reach. Gravity runs expert-built agents that complete an outcome end to end across those tools, so a stronger device layer and a strong cross-app layer can sit side by side rather than compete.

The short version

Sources