Anthropic Gives Its Agent Skills Standard a Vendor-Neutral Home

01Anthropic Donates Agent Skills Standard to Linux Foundation’s AAIF After Adoption by More Than 70 Agent Platforms

Anthropic has donated Agent Skills, an open standard for giving AI agents specialized knowledge and workflows, to the Agentic AI Foundation, or AAIF, under the Linux Foundation. The transfer moves a standard developed by one AI company into a vendor-neutral organization that Anthropic co-founded with Block and OpenAI.

An Agent Skill is a folder containing a SKILL.md file that teaches an AI agent how to perform a particular task or work in a specific domain. The folder can also include scripts and reference materials needed for the task. Anthropic developed the format while building Claude Code, after finding that a general-purpose agent could handle broad knowledge work but often lacked the procedural knowledge required for business tasks.

Developers began publishing thousands of skills on GitHub and building loaders for Codex, Cursor and OpenCode within weeks of Anthropic’s October 2025 launch. Anthropic then published Agent Skills as an open standard in December 2025, with its own GitHub organization and website. The company says more than 70 agent platforms—including ChatGPT, Cursor, GitHub Copilot and VS Code—have since integrated it. It also reports that one million skills have been published and nearly 280 million installations have occurred.

The donation covers the Agent Skills name and logo, its GitHub organization and the agentskills.io domain. AAIF will provide the project’s institutional home, while developers will remain free to use the standard or contribute without becoming AAIF members. Anthropic says the shift is intended to keep the specification open, community-driven and independent of any single vendor.

Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard in their individual capacities. David Soria Parra and Mahesh Murag will serve as lead maintainers. Community members can provide feedback on proposed specification changes through the project’s Discord and participate through GitHub.

Anthropic has proposed an initial roadmap covering versioned releases, a formal process for submitting proposals and security guidance for people who create or install skills. Those measures remain proposals: the announcement does not specify when they will be implemented, how voting rights will be allocated within AAIF or which body will make final decisions about the standard.

02Matrix-Game 4.0 Uses a Separate Game Engine to Constrain Video Generation, Achieving Interactive 480p at 60 FPS on a Single RTX 5090

Matrix-Game 4.0 separates game rules from image generation to tackle a basic problem with playable generative worlds: producing controllable video is not enough. A game must also preserve its state, enforce rules and respond quickly when a player presses a key.

The research team’s new system assigns those jobs to different components. A lightweight, executable game engine maintains the world state and applies interaction rules, while an autoregressive video model—a model that generates each new sequence using earlier output—renders the detailed visuals. The paper reports interactive 480p gameplay at 60 frames per second on a single Nvidia RTX 5090, with 112 milliseconds between a key press and its appearance on screen.

The engine first renders a low-polygon proxy of the game world. That simplified scene gives the video renderer continuous structural guidance, rather than asking it to infer both the game’s mechanics and appearance from generated frames alone. The renderer then converts the proxy into a more detailed visual output, conditioned on dynamic scene captions and memory references.

A multimodal coding agent connects player instructions to the system. It translates players’ ideas into game code and changing scene descriptions, allowing scenes and gameplay to be created or modified through natural language. Because the engine executes the resulting code, mechanics remain enforceable independently of how the video model depicts them.

To build the renderer, the researchers trained it on aligned pairs of low-polygon and high-fidelity videos. They then distilled it into a causal generator that produces four-frame batches using one or two denoising steps, alongside targeted inference optimizations intended to reduce generation time.

In single-step generation, the paper reports steady-state throughput of 83 FPS and peak GPU memory use of 18.2 GiB. Under fixed-prompt tests, throughput reached 67 FPS on an RTX 5080 and 58 FPS on an RTX 5070 Ti. Those figures do not establish comparable stability across complex games, extended sessions or changing prompts, and the source provides no independent reproduction.

The team demonstrated customizable shooting and driving games whose mechanics were enforced by the engine. It also released the code, model weights and low-polygon game assets for researchers and creators.

03AKV Lets AI Agents Manage KV Caches Directly, Cutting Average Cost 27.7% in Paper Experiments

Researchers have proposed AKV, a system that lets long-running AI agents manage their own key-value caches without rewriting their histories. A KV cache stores intermediate data that a large language model reuses when processing earlier tokens, but its storage and input/output demands can become the dominant cost as an agent accumulates observations and actions.

Long trajectories may also exceed a model’s maximum context length. Existing approaches typically trim or summarize the history once it reaches a threshold. That changes the trajectory’s prefix, invalidating later cached states and forcing the model to prefill the retained context again. Keeping less history can reduce that recomputation, but may also hurt task accuracy.

The AKV paper instead proposes “cache engineering”: agents control the active cache while leaving the original trajectory unchanged. Its drop operation excludes selected cached information from future attention. Its repos operation adjusts the position encodings of retained keys, allowing the remaining cache to be repositioned without rebuilding it from a rewritten history.

Direct cache control introduces a serving problem because cache state becomes dependent on the sequence of operations that created it. Identical token sequences can therefore produce different cache states, making ordinary prefix matching insufficient to decide whether a cached result is safe to reuse.

The researchers co-designed the AKV serving engine to handle that distinction. It presents a stateless interface compatible with existing serving APIs, while internally tracking the state needed for cache reuse, eviction and restoration.

In the paper’s experiments, a policy that continually dropped tool responses improved average benchmark scores by 1.7 points and reduced average cost by 27.7% compared with conventional threshold-triggered summarization. With AKV, the experiments reached up to 1.2 times the output throughput and a 1.5-times speedup in P95 request latency. A recursively self-improved cache policy cut costs by another 7.1% relative to an expert-designed policy.

The source does not provide the full model, task and infrastructure configuration for every reported result, and the system has not been independently validated in production.