Apple Core AI Framework: First Look at Apple's Emerging AI Agent Foundation

Apple has published documentation for a new Core AI framework on its developer site. The URL — https://developer.apple.com/documentation/coreai/ — is live and indexed, but the actual content requires JavaScript rendering, so scraping tools and CLI browsers hit a blank page. This is the typical pattern for Apple docs, but the framework name itself suggests a foundational AI layer for apps.
Core AI likely provides on-device inference, model management, and agent-like capabilities for iOS, macOS, and visionOS. Given Apple's focus on privacy and local processing, expect Core AI to enable:
- Running small language models directly on device without network calls
- Integration with Siri and Shortcuts for agentic workflows
- Tokenization and embedding generation for custom AI features
No details are available yet because the documentation page is a JS shell. But the existence of this framework — alongside Apple's open-source MLX machine learning framework and the Core ML model format — signals that Apple is building toward AI agents that run locally, with tight hardware integration.
For now, developers should enable JavaScript in their browser and visit the source page to see if Apple has published technical details. The HN thread (153 points, 22 comments) is largely speculation — the real content has not yet been unveiled.
📖 Read the full source: HN AI Agents
👀 See Also

Trump Pulls Back AI Executive Order Over Fears of Slowing US Tech
President Trump rescinded a Biden-era AI executive order, arguing it could hinder US technological leadership. The move eliminates federal safety reporting mandates.

Rust Will Save Linux from AI: Greg Kroah-Hartman on C Bugs and Rust's Safety Guarantees
Linux stable kernel maintainer Greg Kroah-Hartman says Rust eliminates 60% of kernel bugs at compile time, addressing the flood of AI-discovered CVEs like Dirty Frag and Fragnesia.

Anthropic's DoD Meeting and Chinese AI Labs Distilling Claude
Anthropic's CEO meets with the US Secretary of Defense in what officials describe as a 'shape up or ship out' situation, while the company reports catching three Chinese AI labs conducting massive model distillation of Claude's capabilities.

Google's TimesFM 2.5: 200M-parameter time-series model with 16k context
Google Research released TimesFM 2.5, a 200M-parameter decoder-only foundation model for time-series forecasting with 16k context length and continuous quantile forecasting up to 1k horizon.