Published October 23, 2026 (IST) by the AgentStack News Desk. Facts were checked against the Agentic AI Foundation announcement, Anthropic’s donation post, the AWS Open Source Blog, the Open Responses and OpenSandbox project pages, and our own read of all 10 project repositories on GitHub on October 23.
The Agentic AI Foundation (AAIF) took in four new projects on October 22, 2026: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. The Linux Foundation group, which already hosts the Model Context Protocol (MCP), now runs 10 projects, up from three when it launched in December 2025.
The news came on the first day of AGNTCon + MCPCon North America in San Jose. The four additions cover how agents load skills, how apps call models across providers, how agents are trained in reinforcement learning (RL) environments, and where agent code runs.
Our GitHub count shows the four new projects arrive in very different shape: OpenSandbox logged 1,434 commits in 90 days, while Agent Skills and Open Responses logged seven each. This report follows our editorial standards and adds that count for all 10 Agentic AI Foundation projects.
TL;DR
- What happened: the Agentic AI Foundation added Agent Skills, Open Responses, OpenEnv, and OpenSandbox on October 22, 2026, bringing it to 10 projects.
- Agent Skills: Anthropic donated the SKILL.md standard, its name, GitHub organization, and agentskills.io domain, and says over 70 agent platforms support it.
- Open Responses: OpenAI donated its multi-provider API specification based on the Responses API, with AWS Bedrock, Vercel, and Databricks among the implementers.
- OpenEnv and OpenSandbox: Hugging Face and Alibaba contributed code projects for RL training environments and self-hosted agent sandboxes.
- Our count: OpenSandbox and OpenEnv logged 1,434 and 455 commits in 90 days, while the Agent Skills and Open Responses repositories logged seven each.
- What to do: treat the two specifications as stable formats to build on, and watch their new governance for versioning and security rules.
What the Agentic AI Foundation Announced on October 22
The Agentic AI Foundation announced on October 22, 2026 that four companies had each contributed one project to its neutral governance. Two are specifications, Agent Skills and Open Responses, and two are software projects, OpenEnv and OpenSandbox. They join MCP, Agent2Agent (A2A), AGENTS.md, goose, agentgateway, and Agent Router.
| Project | Contributed by | What it covers | License | GitHub stars (October 23) |
|---|---|---|---|---|
| Agent Skills | Anthropic | SKILL.md folders that give agents task knowledge and scripts | Apache 2.0 | 26,458 |
| Open Responses | OpenAI | One API format for calling models across providers | Apache 2.0 | 1,145 |
| OpenEnv | Hugging Face | Containerized RL environments that speak MCP and Gym | BSD 3-Clause | 2,749 |
| OpenSandbox | Alibaba | Self-hosted sandboxes for coding, browser, and desktop agents | Apache 2.0 | 16,075 |
Manik Surtani, the foundation’s CTO and co-founder, said in the announcement that four different companies chose neutral governance for these projects. The foundation’s projects page now lists all 10 side by side.
Why Anthropic Handed Over Agent Skills
Anthropic donated Agent Skills because it wants the format run like MCP, with input from other vendors on where it goes next. A skill is a folder with a SKILL.md file that teaches an agent one task, plus any scripts or reference files it needs. Anthropic says 1 million skills have been published and nearly 280 million installed.
The handover includes the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain. Engineers from Anthropic, Google, OpenAI, and other companies will maintain the standard as individuals, with David Soria Parra and Mahesh Murag as lead maintainers.
Anthropic says more than 70 agent platforms support the format, including ChatGPT, Cursor, GitHub Copilot, and VS Code. The client showcase file in the agentskills repository listed 46 named clients when we checked, so the public directory trails the adoption figure. The proposed roadmap covers versioned releases, a formal proposal process, and security guidance for people who write and install skills.
What Open Responses Changes for Multi-Provider Agents
Open Responses gives teams one request and streaming format for calling models from different providers, based on OpenAI’s Responses API. Under the foundation, changes to that format go through community governance rather than one vendor’s roadmap. For agent builders, the practical effect is fewer provider-specific adapters to maintain when they switch or mix models.
AWS says on its Open Source Blog that it employs a core maintainer on the project and that Amazon Bedrock has supported the Responses API since December 2025. Bedrock serves that format for models from Amazon, Anthropic, Meta, Mistral, OpenAI, DeepSeek, Google, and others.
An OpenAI spokesperson told The New Stack that Vercel AI Gateway and Databricks Model Serving already implement the specification. Anthropic’s own Messages API is the notable holdout, and many teams still reach Claude through an LLM gateway that translates between formats, the approach compared in our Portkey vs LiteLLM vs OpenRouter review.
OpenEnv and OpenSandbox Bring Running Code
OpenEnv and OpenSandbox are software projects with active codebases, unlike the two specifications. OpenEnv packages RL environments as containerized services that expose MCP for agents and Gym for training libraries. OpenSandbox gives agents isolated environments for running code, commands, browsers, and desktops, from local Docker to Kubernetes clusters.
OpenSandbox reports 64 millisecond median sandbox creation on its Firecracker runtime, per-sandbox egress policies, and a credential vault that lets agents call services without seeing real keys. The credential vault matters after this week’s LiteLLM supply chain compromise, where GitHub tokens left in the project’s own proxy logs led to a poisoned release.
The OpenEnv repository now sits under its own openenv organization. Hugging Face’s Ben Burtenshaw said the committee behind it includes Hugging Face, Microsoft, Reflection, Meta PyTorch, Fleet, Mercor, RadixArk, Modal, and Prime Intellect, which links it to the agent evaluation work covered in our AI agent evaluation tools guide.
What Our GitHub Check of All 10 Projects Shows
We cloned the default branch of all 10 foundation repositories on October 23, 2026 and counted non-merge commits from July 25 to October 23. OpenSandbox led all 10 projects with 1,434 commits from 108 authors. Agent Skills and Open Responses each logged seven, the lowest totals after AGENTS.md, which logged three.
| Project | Commits, 90 days | Authors, 90 days | Top author share | Last commit (UTC) |
|---|---|---|---|---|
| OpenSandbox (new) | 1,434 | 108 | 33.5% | October 23, 2026 |
| agentgateway | 868 | 122 | 54.3% | October 22, 2026 |
| goose | 752 | 146 | 21.4% | October 23, 2026 |
| OpenEnv (new) | 455 | 37 | 57.6% | October 22, 2026 |
| MCP specification | 234 | 46 | 32.0% | October 20, 2026 |
| Agent Router | 198 | 68 | 11.2% | October 22, 2026 |
| A2A | 61 | 35 | 20.0% | October 20, 2026 |
| Agent Skills (new) | 7 | 3 | 71.4% | August 9, 2026 |
| Open Responses (new) | 7 | 1 | 100% | October 22, 2026 |
| AGENTS.md | 3 | 2 | 66.7% | September 10, 2026 |
Every Open Responses commit in the last 90 days came from one OpenAI engineer, Steve Coffey, and the newest one adopted the project’s final technical charter on October 22 (UTC). The Agent Skills repository has not had a commit since August 9, and its specification page last changed on August 4.
Low commit counts are normal for a format that is meant to change slowly, and pull requests can wait on governance. The numbers do show where outside contributors will matter most: both specifications now depend on the foundation to recruit maintainers from other companies.
Company concentration is visible in the code projects too. At least 46.8% of OpenSandbox commits in the window came from Alibaba email addresses, and 62.4% of agentgateway commits came from Solo.io addresses. Those are floor figures, since many contributors commit from personal or GitHub noreply addresses.
What It Means for Teams Building Agents
For teams building agents, the foundation now covers most of the open interfaces an agent touches: tools through MCP, instructions through AGENTS.md and skills, model calls through Open Responses, and execution through OpenSandbox. Our AI agent frameworks guide tracks which frameworks support these formats, and our MCP explainer covers how the protocol fits in.
Neutral governance does not make an instruction format safe. Forever Security reported on October 22 that it found a malicious AGENTS.md in a customer’s Downloads folder that told Codex to install a malicious npm package, and it reproduced a similar attack on Claude Code. Skills carry the same risk, because a SKILL.md can bundle scripts an agent will run.
Anthropic’s proposed roadmap lists security guidance for skill authors and installers, which would be the first formal rule set for the format. Until it ships, the controls in our MCP security tools guide and an MCP gateway policy layer, such as the Red Hat MCP Gateway we covered, are the closest available checks on what agents load.
Four Steps to Take This Quarter
These steps follow from what the foundation announced and from our repository check. They apply to any team that ships agents using skills, multiple model providers, or self-hosted sandboxes, whichever framework sits on top, and none of them requires waiting for the new governance to finish its first release.
- Inventory skills: list every SKILL.md folder and AGENTS.md file your agents load, and review any bundled scripts the same way you review dependencies.
- Pin a model format: if you call several providers, test the Open Responses compliance suite against your gateway before adopting it as the internal standard.
- Trial a sandbox: run OpenSandbox’s egress policies and credential vault in a staging cluster before giving coding agents network access.
- Watch governance: follow the Agent Skills and Open Responses repositories for versioned releases, since both have had little merged activity since July.
Frequently Asked Questions
What did the Agentic AI Foundation announce on October 22, 2026?
The Agentic AI Foundation added four projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. The additions bring the Linux Foundation group to 10 projects, alongside MCP, A2A, AGENTS.md, goose, agentgateway, and Agent Router. It announced them at AGNTCon + MCPCon North America in San Jose.
What are Agent Skills?
Agent Skills is an open format for packaging task knowledge for AI agents. Each skill is a folder with a SKILL.md file holding a name, a description, and instructions, plus optional scripts and reference files. Agents load a skill when a task needs it. Anthropic says more than 70 platforms support the format today.
What is Open Responses?
Open Responses is an open API specification based on OpenAI’s Responses API. It defines one schema for requests, streaming events, and tool calls so the same code can call models from different providers. Amazon Bedrock, Vercel AI Gateway, and Databricks Model Serving support it, while Anthropic still uses its own Messages API.
Who controls Agent Skills after the donation?
The Agentic AI Foundation now owns the Agent Skills name, logo, GitHub organization, and agentskills.io domain. Engineers from Anthropic, Google, OpenAI, and other companies maintain it as individuals, led by David Soria Parra and Mahesh Murag. Anyone can keep using the format or contribute without joining the foundation, according to Anthropic.
Is OpenSandbox an alternative to hosted agent sandboxes?
OpenSandbox is a self-hosted option. It runs agent sandboxes locally on Docker or on Kubernetes, with SDKs in five languages, egress policies, and a credential vault. Hosted services handle the infrastructure for you, so the choice comes down to whether your team wants to operate its own sandbox cluster.
The Bottom Line
The Agentic AI Foundation now governs the main open formats an agent uses, from tools and skills to model calls and sandboxes. The two specifications it took in this week are widely adopted but thinly maintained on GitHub, while the two code projects arrive busy. The next test is whether other companies start committing to Agent Skills and Open Responses.
We will update this story when either specification publishes a versioned release or its first security guidance. Send corrections through our contact page.
Related coverage: Agent frameworks, Claude Managed Agents dynamic workflows, Red Hat MCP Gateway, LiteLLM supply chain compromise, and LLM gateways.