● AGENTS · PLATFORM

Claude's agent stack graduates to general availability

Computer use, the browser tool, the Skills API and the Files API all leave preview on the Claude Platform — a direct move to automate the long tail of work that has no API to call.

Anthropic crossed a threshold this week: four agent primitives — computer use, the browser tool, the Skills API and the Files API — are now generally available on the Claude Platform. Together they target work that lives outside clean API boundaries: legacy line-of-business apps, browser-only dashboards and the manual drudgery of moving data between screens.

The pitch is fewer round trips per task. A computer-use agent can drive a real screen instead of waiting for an integration to exist, while the browser tool reaches applications that live only on the web. Versioned skills and reusable files give Claude Managed Agents a stable substrate — teams can pin a capability, share files across runs, and build on a stack that behaves the same in production as it does in staging. For a lab that has spent the year casting agents as the next interface, this is the moment the scaffolding becomes a product.

By the FAV0 desk · Anthropic / Claude Platform

● INFRASTRUCTURE

First NVIDIA Vera Rubin racks power OpenAI's training stack

The first NVIDIA Vera Rubin racks are now running OpenAI's training stack — a milestone in the accelerated-computing roadmap behind frontier AI. NVIDIA called the deployment the start of a new infrastructure generation.

● INFERENCE

NVIDIA's NeMo Switchyard routes each agent step to the right model

The right model depends on the task. NeMo Switchyard lets developers route every step of an agent workflow across a chosen model pool by their own quality, latency and cost criteria. Kari Briski argues agent workflows need model routing because no single model wins every step.

● EVALUATION

Meta previews WildArtifactBench for real-world agent evals

Meta is previewing an internal evaluation framework that scores agents on complex, real-world tasks across diverse deliverable formats — using win rates and Elo scores from human and agentic preference judges rather than strict ground-truth.

Muse Spark 1.2 turns visuals into working code

Meta's Muse Spark 1.2 spans a broad set of multimodal tasks, from turning visuals into working code to translating perception into physical action. It also brings audio-visual understanding to the video-heavy workflows common in real-world enterprise use.