Apple WWDC 2026: Xcode 27 Agent Mode and Foundation Models Get a Major Developer Upgrade

Agent Mode turns Xcode into an autonomous coding loop. Xcode 27 introduces Agent Mode, which integrates coding agents — from Anthropic, Google, and OpenAI — directly into the development workflow. Agents get interactive planning, multiturn Q&A, and a canvas that renders Markdown alongside code changes and previews. Critically, agents can validate their own work: they write and run tests, use Playgrounds, check previews, and interact with the iOS simulator via a new Device Hub. The result is a closed feedback loop where the agent proposes a fix, exercises it against a running simulator, and iterates — without the developer manually switching contexts.

This is direct competition for AI coding tools like Cursor and GitHub Copilot Workspace that have been building agentic coding loops in the browser or VS Code. Apple's version is native to Xcode, has first-party simulator access, and ships the Anthropic, Google, and OpenAI agent integrations as built-in options — no third-party extensions required. Apple notes support for Model Context Protocol (MCP) and a new Agent Client Protocol, so custom integrations are also on the table.

Foundation Models gains image input and a model abstraction layer. The on-device Foundation Models framework now supports vision — developers can attach UIImage, NSImage, CGImage, Core Image, or CoreVideo pixel buffers directly in a session prompt. The framework also introduces a new `LanguageModel` protocol that lets local and server models back the same `LanguageModelSession`, meaning you can swap between on-device, Private Cloud Compute, and third-party frontier models (Anthropic and Google ship Swift packages for this) without rewriting your inference code. Private Cloud Compute sessions now expose a 32K context window with configurable reasoning levels (light or deep) and per-token usage tracking including cached and reasoning tokens.

Core AI framework: run full LLMs locally on Apple Silicon. New at WWDC is a Core AI framework — Apple's description is "the best way to run models on device," optimized for Apple Silicon's unified memory and Neural Engine to support full-scale, open-source LLMs. Two implementations ship with it: `CoreAILanguageModel` and `MLXLanguageModel` for local open-source models. The framework's core layer is open-sourced and runs wherever Swift runs, including Linux servers, giving teams the option to use the same model abstraction layer server-side.

The `fm` CLI brings on-device models to the terminal. macOS 27 ships an `fm` command-line tool that exposes both on-device and Private Cloud Compute models without any account setup, API keys, or auth. `fm chat` opens an interactive session; piping into `fm` enables summarization, extraction, or content generation in shell scripts. For developers building data pipelines or automation, this is a meaningful addition — on-device inference from the terminal, private by default.

A new Python SDK targets data scientists and researchers. The Foundation Models framework now ships a Python SDK (`apple_fm_sdk`) that exposes the same on-device model as the Swift API. It covers model availability checks, session creation, and structured response generation. This is aimed squarely at the Python ML workflow — data scientists and researchers who want to evaluate or prototype against Apple's on-device models without switching to Swift.

Small Business Program developers get free Private Cloud Compute access. Apple's press release states that developers with fewer than 2 million total first-time App Store downloads can access the next generation of Apple Foundation Models running on Private Cloud Compute at no cloud API cost. Apple has not published a per-token quota, rate limits, or what happens when usage exceeds any unpublished threshold — so treat this as a free tier with unspecified limits until Apple's developer documentation is more explicit.

New system tools round out the release. The framework ships vision-backed system tools — `BarcodeReaderTool` and `OCRTool` — and a Spotlight-powered search tool that enables fully local RAG (retrieval-augmented generation) without a third-party vector database. A new Swift Evaluations framework lets teams measure and quantify the accuracy of their intelligence features, which matters when agent-assisted features ship to production users.