Apple’s New Siri Public Beta Redefines Mobile Assistants with LLM‑Based Agent Architecture

Preface

In July 2026 Apple opened a public beta of a completely re‑engineered Siri. Unlike a routine feature update, the overhaul replaces the long‑standing intent classification pipeline with a large language model (LLM)‑centered agent architecture that supports cross‑app actions, continuous dialogue memory, and real‑time multimodal understanding. Competing assistants such as Google Gemini and Samsung Galaxy AI are also advancing, indicating a paradigm shift in mobile assistants.

1. Technical Bottlenecks of the Legacy Siri

Siri’s original architecture, unchanged since its 2011 launch, follows a classic NLU pipeline: ASR → NLU → Intent Router → Domain Fulfillment. Adding new capabilities required manually defining intent templates and slots, leading to high extension costs and poor handling of out‑of‑distribution utterances.

Deeper issues stem from tight coupling between domain modules (music, navigation, messaging). Requests that span domains, e.g., “share the song I just played with Zhang San,” cannot be processed because modules do not share context. Multi‑turn state management is fragile, losing context with slight deviations from predefined flows. Apple’s 2024 Apple Intelligence introduced patch‑style fixes, but a fundamental redesign only materialized in the 2026 beta.

2. New Siri Agent Architecture

The core change is an LLM‑driven agent design described as “one central brain + a set of callable tools.”

Chain‑of‑thought reasoning and tool invocation. Siri adopts a ReAct‑like pattern where the model alternates between “thinking” and “acting.” For example, when a user says “Check tomorrow’s weather in Beijing; if it rains, cancel my outdoor plans,” the model first calls a weather API, then decides based on the result whether to invoke a calendar cancellation.

Persistent dialogue memory. A hierarchical memory system stores short‑term (current session), mid‑term (recent interaction summaries), and long‑term (user preferences) data. Long‑term memory resides in an encrypted on‑device database and is retrieved via vector search to enrich each conversation.

Native multimodal support. The perception layer is a unified multimodal encoder that processes speech, screenshots, and camera input together, enabling commands like “translate this” while pointing at on‑screen content.

3. Edge‑Cloud Collaborative Inference

Balancing performance and privacy is the most challenging engineering problem. Apple implements a three‑level inference hierarchy.

Level 1 runs on the A20/M5 Neural Engine with a ~3 billion‑parameter model, quantized and distilled to respond within 200 ms for simple tasks such as setting alarms or launching apps, requiring no network connection.

Level 2 is the upgraded Private Cloud Compute (PCC) introduced in 2024. User data is processed inside Apple’s Secure Enclave and discarded after use; the server never retains raw data. The 2026 PCC adds inference acceleration and model hot‑switching for moderately complex multi‑step tasks.

Level 3 is the full cloud‑side Apple Foundation Model (AFM), now in its second generation, handling the most complex reasoning, code generation, and long‑form planning.

Routing decisions are made by a lightweight on‑device classifier that evaluates task complexity, privacy sensitivity, and network quality, dynamically selecting the appropriate level. The classifier continuously learns by observing actual completion quality at each level.

4. Privacy Computing and Security Mechanisms

Privacy remains Apple’s core differentiator, and the new Siri invests heavily in engineering safeguards.

Device‑first principle. Processing of contacts, photos, health data, and location defaults to on‑device execution. Only when the local model cannot handle a request does the system fall back to the cloud, after applying field‑level de‑identification.

Verifiable PCC transparency. Every software update to a PCC node is published to a public Transparency Log, allowing security researchers to download and audit the exact system image, providing provable consistency between hardware‑enforced code and the logged version.

Differential privacy and federated learning. User interaction data used for model improvement is first perturbed with local differential privacy noise before leaving the device. Long‑term preference learning occurs via federated learning, uploading only gradient updates rather than raw data.

5. Developer Ecosystem: From App Intents to Agent Actions

For third‑party developers, the shift means a fundamentally new integration model. The legacy App Intents framework required manual definition of intent schemas, parameter types, and dialogue flows for each capability, resulting in high effort and low flexibility.

The new Agent Actions framework lets developers describe what their app can do (capability description) and the required parameters and outcomes (interface contract) in a structured format. The LLM agent automatically understands these descriptions and invokes the appropriate actions at runtime.

For example, a food‑delivery app declares three actions—place order, query order, cancel order—along with their parameter schemas. Siri can then handle any natural‑language request related to those actions without the developer enumerating every possible utterance.

Apple also provides a SiriKit Playground where developers can simulate Agent Action calls, test edge cases, and explore multi‑app coordination scenarios.

6. Industry Landscape and Outlook

The new Siri release reflects broader trends in the mobile assistant arena.

From command execution to task agency. Traditional assistants followed a one‑way command‑execute model. Next‑generation assistants now plan and decide, accepting high‑level goals and decomposing them into executable steps. Google’s Gemini on Android 16 shows similar cross‑app capabilities and proactive suggestions.

On‑device models become a necessity. With Qualcomm Snapdragon 8 Elite and MediaTek Dimensity 9400 delivering stronger NPU performance, running 3‑billion‑parameter (or even 7‑billion) models on‑device is feasible, enabling offline functionality and stronger privacy guarantees.

Multimodal understanding as a baseline. Siri, Gemini, and Samsung Galaxy AI all converge on unified visual‑speech‑text perception, turning interactions like “translate this” while pointing at the screen into everyday experiences.

Apple’s execution speed in this architectural overhaul exceeds many expectations, challenging the narrative that Siri has lagged behind. Nevertheless, the public‑beta phase still requires large‑scale user validation. Key observation points include the reliability of agent‑driven complex task chains, third‑party app onboarding speed, and edge‑cloud routing performance under weak network conditions.

The rising technical bar for mobile assistants ultimately benefits users, even if it raises development complexity.

Based on WWDC 26 technical documentation and hands‑on experience with the public beta; some architectural details may change before the final release.