On October 22, 2026, at AGNTCon + MCPCon North America in San Jose, the Agentic AI Foundation (AAIF) announced four new hosted projects: Agent Skills, contributed by Anthropic; Open Responses, contributed by OpenAI; OpenEnv, contributed by Hugging Face; and OpenSandbox, contributed by Alibaba. The additions bring AAIF to ten hosted projects in its first year — up from the three it launched with in December 2025. The Linux Foundation also announced the same day that AAIF has surpassed 300 members: 305 organizations, with 60 joining this quarter alone, including new Gold Members Ant Group, the Gates Foundation, MongoDB, Nutanix, Royal Bank of Canada, Vercel and Vultr.
The headline donation is the one with the largest installed base. Anthropic is handing over Agent Skills, the open standard it launched roughly a year ago for packaging instructions, scripts and resources that an AI agent can load on demand — a way to give agents new capabilities and domain expertise without retraining or rebuilding them. Anthropic isn't just contributing the code: it is transferring the Agent Skills name and domain to the foundation (per reports of the October 22 announcement). The specification is open source under Apache 2.0 and CC BY 4.0.
The numbers behind the donation
| Agent Skills, as of October 22, 2026 | Figure |
|---|---|
| Agent platforms that already support the standard | 70+ |
| Skills published | Over 1 million |
| Total installations | Nearly 280 million |
| Time for the skills.sh registry to reach 1 million skills | 7 months |
| Licensing | Apache 2.0 and CC BY 4.0 |
| Governance after the donation | Technical Steering Committee; core maintainers from multiple organizations, serving as individuals |
A million published skills in seven months, used across software engineering, marketing, design, writing and research, is not a hobby ecosystem — it is infrastructure. And infrastructure built on a single company's goodwill is exactly what foundations exist to de-risk. As David Soria Parra, co-creator and lead maintainer of MCP at Anthropic, put it: Anthropic launched Agent Skills to give developers and enterprises an open standard for making agents more capable, and is donating it so that anyone building with skills can shape what comes next.
The MCP playbook, run again
This is the second time Anthropic has made this exact move. On December 9, 2025, it donated the Model Context Protocol — the connective tissue that lets agents reach tools and data, now running in production at companies like Uber with 800 servers and 5,000 tools — to the newly created Agentic AI Foundation, co-founded with Block and OpenAI. MCP went from a proprietary connector spec to a foundation-governed protocol, and adoption went vertical: it is now the de facto way agents talk to the world, celebrated this week at a conference named after it.
The strategic logic is not altruism. Donating a standard removes the biggest objection to adopting it: that a single vendor can change the terms, the pricing, or the roadmap and strand everyone downstream. The vendor keeps influence through maintainership; the ecosystem gets a neutral home. Under AAIF, Agent Skills will be guided by a Technical Steering Committee with core maintainers drawn from multiple organizations who serve as individuals — the same governance pattern that made Linux credible. It is the same open-governance shift we have watched remake agent sprawl tooling like WSO2's Agent Manager, and it mirrors why Uber bet its production stack on MCP while staying vendor-neutral.
The four new projects, by layer
What makes the October 22 batch interesting is that the four projects cover four different layers of the agentic stack:
| Project | Contributor | Layer of the agentic stack | What it does |
|---|---|---|---|
| Agent Skills | Anthropic | How agents gain skills | Open format packaging instructions, scripts and resources agents load on demand |
| Open Responses | OpenAI | How applications talk to models | Shared schema and streaming events for multi-provider, interoperable LLM interfaces, based on the OpenAI Responses API; already exposed by Databricks Model Serving and Vercel AI Gateway |
| OpenEnv | Hugging Face | How agents are trained and evaluated | Framework and community hub for reinforcement-learning environments; Turing's Calendar Gym reported roughly 90% task success with explicit identifiers versus roughly 40% with natural-language task descriptions |
| OpenSandbox | Alibaba | Where agents run | Apache-2.0 sandbox platform with lifecycle and execution APIs over Docker and Kubernetes runtimes; an execd daemon runs inside each sandbox |
Manik Surtani, CTO and co-founder of AAIF, framed the milestone: "A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10." The four projects join Agent2Agent (A2A), Agent Router, agentgateway, AGENTS.md, goose and MCP under the foundation's open, vendor-neutral governance.
OpenSandbox's contributor made the point explicitly: "While OpenSandbox was created at Alibaba, we don't want its future to be defined by a single company," said Xiaobin Xu, head of development infrastructure at Alibaba. Similarly, OpenEnv's maintainers describe a committee spanning Hugging Face, Microsoft, Meta-PyTorch and others — a multi-company Technical Committee submitting the specification to a vendor-neutral home through a public AAIF proposal.
Why a Canadian perspective matters here
Buried in the Gold Member list is a detail we track closely: the Royal Bank of Canada is now a Gold Member of AAIF. A major Canadian bank joining the foundation that governs the infrastructure layer of agentic AI — the same week as the Axios report on unvetted agents inside U.S. bank accounts — tells you where financial institutions think the real risk-management work will happen. You do not fix the agent-access problem by blocking agents one by one; you help govern the protocols, identity and sandbox standards they are built on. RBC's seat at AAIF, alongside a global philanthropy and infrastructure providers, is a bet that standards, not firewalls, decide how agent governance matures.
What your business should actually check when adopting a standard
Whether you run a two-person operation or a serious enterprise stack, the foundation era of agent tooling changes your due diligence:
- Check who governs, not who created. Agent Skills under AAIF has a multi-organization Technical Steering Committee. A skill format under one company has a roadmap. Weight the first; discount the second.
- Verify licensing continuity. Apache 2.0 and CC BY 4.0 for Agent Skills means your investments in skills survive corporate strategy changes. Proprietary formats offer no such promise.
- Prefer interoperable layers. Open Responses defines a provider-agnostic API surface; Databricks and Vercel already expose it. Lock-in is now a choice, not a default.
- Where agents run matters more than where they are written. OpenSandbox codifies sandboxed execution over Docker and Kubernetes — the same isolation thinking behind the agent-security wave we have covered from Straiker to Cohere's AISLE acquisition.
- Watch the governance layer of everything. Agent identity, registries and runtime gateways — the territory of Workday's C1.ai acquisition, ServiceNow's AI Passport, and the inherited-agent problem — is being standardized as we speak. The organizations at the table now write the rules your agents will live under.
If you are choosing agent infrastructure for your business in 2026, this is the layer of decisions most teams skip — then pay for at migration time. Get in touch if you want a vendor-neutral review of your agent stack.
Read next:
