Apple shipped Siri AI and the Core AI framework at WWDC 2026 — the first-party developer API for on-device LLM features. Anthropic's Cat Wu and a wave of AI developers are now "ditching" prompting and writing loops — small systems that run the agent, evaluate the output, and re-prompt until convergence.

Siri AI, Apple Core AI, loop engineering

Hey guys, Mr. Technology here — let me break this one down.

What You Need to Know: Apple shipped "Siri AI" at WWDC 2026 alongside a brand-new Core AI framework that replaces Core ML and exposes Apple's on-device foundation models to developers. Anthropic's Cat Wu — the leader who built Claude Code — said he "ditched" prompting agents and now designs loops that prompt themselves. The Anthropic/Apple convergence on "system, not prompt" is the most important developer pattern shift of 2026.

Why It Matters

Apple's Core AI framework is the new developer API for the foundation-model era. Anyone shipping a Mac/iOS app that wants to do on-device LLM inference now has a first-party path. The same goes for the new Siri app-integration hooks.
"Loop engineering" replaces "prompting" as the high-leverage skill. The shift: instead of writing the prompt that tells the agent what to do, you write the small system that loops, evaluates, and re-prompts. Cat Wu's "ditched prompting" comment is the canary.
The two stories together mean: the model is the cost of entry, the system is the product. Apple's Core AI gives you the model. Loop engineering gives you the system. Builders who only know how to write prompts are about to be outperformed by builders who know how to write loops.

Apple announces Siri AI at WWDC 2026

At WWDC 2026, Apple shipped Siri AI — a rebrand of Siri as a more conversational LLM-powered assistant with Google-influenced updates to its on-device Foundation Models, available across iOS 27, iPadOS 27, macOS 27, and visionOS 27 this fall (Apple Newsroom, 6/8/2026; Ars Technica, 6/8/2026). Siri can now handle multi-step personal tasks like researching concert tickets end-to-end, with the architecture now leaning on Google's foundation-model research alongside Apple's own in-house teams.

Craig Federighi's leadership page now explicitly lists "Apple's core AI efforts, including the development of foundation models, applied AI technologies, and the research that powers intelligent experiences" as his remit (Apple Leadership). For a builder, the practical shift: the new Siri has hooks for third-party app integration, and the LLM behind it can now plan multi-step tasks across apps — meaning the apps you expose to Siri become meaningfully more discoverable starting in iOS 27.

Apple ships the Core AI framework, replacing Core ML

Alongside the consumer Siri announcement, Apple introduced the Core AI framework — the foundation-model-native successor to Core ML. The framework exposes Apple's on-device foundation models, the Foundation Models API, and integration hooks for app developers (CNET, 6/2026; Apple Magazine, 3/2026). The pitch: developers can build AI features that run locally, respect Apple's privacy posture, and don't require shipping user data to a third-party model provider. The first-party Foundation Models repo and integration give Mac/iOS apps a true on-device path for LLM features.

The Siri rebuild is the visible product. Core AI is the developer surface. For a builder who has been waiting for first-party Apple LLM APIs before committing to on-device AI features, the wait is over — Core AI is available in the iOS 27 beta now, and the public API will be stable in the fall. Anything you built on Core ML needs a migration plan.

Loop engineering replaces prompting

The New Stack published a piece on loop engineering — the practice of designing the system that prompts a coding agent instead of prompting the agent yourself (The New Stack). Firecrawl's blog framed the same idea more concretely: "Loop engineering is the practice of designing the system that prompts a coding agent instead of prompting the agent yourself. You write a small system that runs the agent in a loop, evaluates its output, and re-prompts" (Firecrawl).

The pattern got its name when Anthropic's Cat Wu — the leader who built Claude Code — said publicly that he now "ditches" prompting and designs loops instead. The argument: a single prompt is a static guess. A loop is an evolving system that uses evaluation signals to converge on a result. Linas's Substack piece captures the practical guide: "Stop prompting AI agents and design the systems that prompt them" (Linas's Newsletter).

The Twitter conversation that pushed "loop engineering" into the trending list is the clearest distillation: "Write loops, set a goal, and let the agent run until it reaches it" (LinkedIn — Michael R Larson).

The pattern looks like:

A small driver loop that calls the agent with a goal
An evaluator that scores the output (compile, tests, lint, custom metric)
A re-prompt that uses the eval signal to refine the next attempt
A termination condition (success, budget, max iterations)

The value of a loop over a prompt: it's reproducible, it's evaluable, and it doesn't depend on the model being "lucky" on a single inference. The cost: you have to write the eval. If your eval is "looks right to me," you have a loop that confidently produces the wrong answer 1,000 times in a row.

The Take

Three stories, one through-line: the model is now the cost of entry, and the system is the product.

Siri AI is the consumer UX layer that finally ships. Core AI is the developer surface that makes on-device LLM features real. Loop engineering is the discipline that determines whether the agent you ship on top of either one converges or hallucinates. A builder who can write a Core AI integration but only knows how to one-shot prompt is going to ship a feature that works in the keynote demo and breaks the first time the user asks something novel. A builder who writes a loop with a real eval is going to ship a feature that converges.

The 2027 hiring market for AI engineers is going to look like:

Model-evaluator engineers (write the loop, the eval, the re-prompt)
System integrators (wire Core AI / Claude / Fable into the existing stack)
Eval-data curators (the maintainer role that keeps Anthropic's 95% from decaying to 65%)

The prompt-only engineers are about to be the equivalent of 2010 web designers who only knew HTML. The skill is moving up the stack. The system is the new prompt.

Quick Summary

Apple shipped Siri AI (the consumer LLM assistant) and the Core AI framework (the developer surface for on-device LLM features in iOS 27 / macOS 27) at WWDC 2026. Cat Wu (Claude Code lead) and a wave of AI developers are now "ditching" prompting and writing loops — small systems that run the agent, evaluate the output, and re-prompt until convergence. The model is the cost of entry. The system is the product.

Sources:

Siri AI , Apple Core AI , loop engineering