Agent Skills, Open Responses, OpenEnv and OpenSandbox: The New AI Agent Standards Explained

On October 22, 2026, the Linux Foundation's Agentic AI Foundation (AAIF) took on four new open projects: Anthropic's Agent Skills, OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox. That brings AAIF to 10 hosted projects, up from three at its December 2025 launch. Together the four cover what an agent knows, how it talks to models, how it gets trained and where it can safely run code.

What did the Agentic AI Foundation announce?

AAIF said it was adding four projects from four different companies, all under what it calls open, vendor-neutral governance. The Linux Foundation press release came out at AGNTCon + MCPCon North America, held October 22-23 at the San Jose McEnery Convention Center.

The foundation launched in December 2025 with three projects: the Model Context Protocol (MCP), goose and AGENTS.md. It has since added the Agent2Agent (A2A) protocol, Agent Router and agentgateway. Manik Surtani, AAIF's CTO and co-founder, said the four contributions came from different companies so that no single company controls the outcome.

What is Agent Skills, in plain English?

Agent Skills is a standard way to package know-how so any compatible AI agent can load it when a task calls for it. In Anthropic's announcement, a skill is a folder with a SKILL.md file that teaches an agent a task or domain, plus any scripts or reference files it needs.

Think of it as an onboarding binder for an agent. Anthropic says it built skills because Claude Code was good at general knowledge work but lacked the step-by-step procedures many business tasks need. It published the standard in December 2025.

• Adoption claims: Anthropic says more than 70 agent platforms have integrated Agent Skills, naming ChatGPT, Cursor, GitHub Copilot and VS Code. It also says about 1 million skills have been published and nearly 280 million installed. These are Anthropic's own figures, which The Neuron also repeated from a Claude Developers post on X.
• What moves to AAIF: the Agent Skills name and logo, the GitHub organization and the agentskills.io domain.
• Who maintains it: engineers from Anthropic, Google, OpenAI and other companies, acting as individuals. David Soria Parra and Mahesh Murag are lead maintainers.
• Roadmap: versioned releases, a formal proposal process and security guidance for people who write and install skills.

What is Open Responses?

Open Responses is a shared specification for how apps send requests to large language models, so the same code can run against different providers. It is based on OpenAI's Responses API. The press release says it defines a common schema, consistent streaming events and extensible tooling.

The New Stack reports that the spec standardizes building blocks such as messages and tool calls. Its listed supporters include AWS Bedrock, Databricks, Hugging Face, Nvidia, Ollama, OpenRouter, Red Hat, Vercel and vLLM. An OpenAI spokesperson told the publication that Vercel AI Gateway and Databricks Model Serving already implement it. In plain terms, it is an attempt at a universal plug for LLM APIs.

What is OpenEnv?

OpenEnv is a framework and community hub for building the practice environments where AI agents learn through reinforcement learning (RL). Each environment is a containerized service that exposes a familiar agent interface: MCP plus Gym, the protocol RL training libraries use.

Hugging Face and Meta's PyTorch team first launched the OpenEnv hub in October 2025. Hugging Face's Ben Burtenshaw said that democratizing AI "requires a fully open stack." Per the press release, the OpenEnv committee includes Hugging Face, Microsoft, Reflection, Meta-PyTorch, Fleet, Mercor, RadixArk, Modal and Prime Intellect, among others.

What is OpenSandbox?

OpenSandbox is Alibaba's platform for giving AI agents an isolated place to run code, so a mistake or misuse doesn't reach production systems. The OpenSandbox GitHub repository describes isolated environments where agents can run commands, manage files and operate browsers or desktops, using Docker locally or Kubernetes at scale. It is licensed under Apache 2.0.

The press release lists use cases including coding agents, GUI agents, agent evaluation and RL training. A sandbox protocol lets teams plug in their own runtimes. Alibaba's Xiaobin Xu said the company does not want the project's future "to be defined by a single company."

What are critics saying?

Skeptics raise three main concerns: who really controls the foundation, whether the standards end fragmentation, and security.

1. Vendor influence. The Register ran the news under the headline "Vendor-affiliated AI agent tech gets adopted by friendly, funded foundation." It pointed out that the projects come from companies with deep ties to the foundation. The same report says AAIF has more than 300 members; the press release gives no member count.
2. Fragmentation hasn't gone away. The New Stack notes that Anthropic is not on the Open Responses supporter list because it uses its own Messages API. Google has its own Interactions API, and several providers expose Anthropic-compatible endpoints. The publication also notes that the project's contributor guide says its core schema files are copied from OpenAI's API.
3. Skills security. The skills format is now governed, but public skill marketplaces are not. In February 2026, Snyk researchers analyzed 3,984 skills and confirmed 76 as malicious. ESET's H1 2026 threat report, as reported by Expert Insights, flagged more than 3,000 of roughly 900,000 skills as directly malicious. ESET advises treating skills like any other third-party software.

Anthropic's roadmap includes security guidance for skills, but it has not described that guidance in detail.

What does an agent-first web mean for brands?

Shared standards make it easier for one agent to work across many tools, which makes it more important for brands to show up accurately wherever those agents look. When MCP connects agents to data, skills package procedures and Open Responses lets the same agent switch models, a buyer's research may run through several models and tools before a person reads anything.

That changes the work of generative engine optimization (GEO). Ranking pages for humans is no longer enough. Brands also need content and data that agents can parse, cite and trust:

• Clear, consistent facts about products, pricing and policies that any model can quote without guessing.
• Structured, current sources, so agents don't fill gaps with outdated third-party claims.
• Regular checks on how different models describe and recommend your brand, because the model behind an agent may change from task to task.

Tools like MarkGrid's Model Share are built to track how often, and how, a brand appears across AI models. For teams still mapping where agents fit into the funnel, our explainer on what AI agents in marketing are is a practical starting point. The standards are still settling, but planning for agents as a real audience is already a reasonable bet.

FAQ

What is the Agentic AI Foundation?

The Agentic AI Foundation (AAIF) is a Linux Foundation body launched in December 2025 to host open standards and tools for AI agents. It now hosts 10 projects, including MCP, A2A and AGENTS.md.

Who contributed the four new AAIF projects?

Anthropic contributed Agent Skills, OpenAI contributed Open Responses, Hugging Face contributed OpenEnv and Alibaba contributed OpenSandbox. AAIF announced them on October 22, 2026.

Is Agent Skills the same as MCP?

No. MCP connects agents to external tools and data. Agent Skills packages instructions, scripts and resources that teach an agent how to do a task. The two are complementary, and both are now AAIF projects.

Does Anthropic support Open Responses?

The New Stack reported that Anthropic was not on the launch list of Open Responses supporters, noting that it maintains its own Messages API. The publication speculated that Anthropic could adopt the spec later.

Are AI agent skills safe to install?

Not automatically. Security researchers at Snyk and ESET have found malicious skills in public marketplaces. Vet skills before installing them and treat them like any other third-party software.

Why should marketers care about agent standards?

Shared standards let agents work across many tools and models, so more product research and buying decisions may pass through AI agents. Brands need accurate, consistent information that those agents can find and cite.

Last updated October 24, 2026.