📰 AI Daily Recap

A quiet Saturday — Latent Space's own roundup was literally titled "not much happened today" — but two platform stories are worth a platform engineer's attention. At WWDC 26, Apple introduced Core AI, the official successor to Core ML, a framework for running LLMs and generative models entirely on-device against Apple Silicon. And on the terms-of-service front, InfoQ reports that using Claude Fable 5 (or Mythos 5) on Amazon Bedrock now requires opting into provider data-sharing — prompts and outputs flow to Anthropic for 30-day retention with human review, a break from how earlier Bedrock models kept inference data isolated.

The rest of the day was Show HN agent-tooling: a wave of small, single-purpose utilities aimed at the same workflow most of us already live in. Persona.js open-sourced a framework-free, WebMCP-native agent UI library; Logslim compacts test/build output before an agent reads it; taste-ai packs per-session "taste" context; and a name.com CLI gives Claude Code and Codex a skill to manage DNS. The connective tissue is clear: the tooling layer is consolidating around agents as the primary user.

If there's a thread worth tracking, it's agent security and isolation — Cordium (a self-hosted, identity-based sandbox built on Kubernetes) and an industrial middleware that fronts PLC data behind REST/gRPC both lean into keeping infra secrets away from both developers and the AI agents acting on their behalf.