The Agentic AI Foundation has grown to 10 hosted projects in its first year after adding four new projects at AGNTCon and MCPCon North America: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. The foundation launched in December 2025 with three projects and now covers agent capabilities, model interoperability, reinforcement learning environments and safe execution under one vendor-neutral governance model.
That expansion matters because agentic AI is developing as a stack rather than a single product category. Systems need ways to acquire specialized knowledge, communicate with different models, train against consistent environments and execute code with appropriate isolation. The new projects address those layers individually, but their placement inside the same foundation gives them a common governance path.
Four New Projects Fill Gaps Across the Agent Stack
Agent Skills packages instructions, scripts and reference resources so agents can load domain expertise on demand. A skill can include a SKILL.md file along with supporting material, allowing procedural knowledge to travel between agent products instead of remaining hard-coded inside one platform.
Anthropic reported that more than 70 agent platforms had integrated the format by October 2026, including ChatGPT, Cursor, GitHub Copilot and Visual Studio Code. The company also said developers had published about 1 million skills and generated nearly 280 million installations since publishing the standard in December 2025. Those figures suggest an audience well beyond the coding assistant market, even though the format originated during work on Claude Code.
Open Responses takes a different layer. Built from the OpenAI Responses API, the specification defines a shared request schema, consistent streaming events and extensible tooling for interfaces that can work across model providers. The practical goal is to let developers describe an application once and deploy it with limited translation work when models or providers change.
OpenEnv addresses the training and evaluation layer. It provides a framework and community hub for building reinforcement learning environments through a containerized service that exposes a familiar agent interface, Model Context Protocol support and the Gym training protocol. Its design goals include interoperability, easier sharing and scalability.
OpenSandbox focuses on where agent workloads run. The platform offers software development kits for multiple languages, unified sandbox APIs, and Docker and Kubernetes runtimes. Its stated use cases include coding agents, graphical interface agents, agent evaluation, code execution and reinforcement learning training. A defined sandbox protocol sets out lifecycle and execution interfaces so teams can add custom runtimes.
Why the Agentic AI Foundation Model Matters
The new projects join six existing foundation initiatives: the Agent2Agent protocol, Agent Router, agentgateway, AGENTS.md, goose and the Model Context Protocol. Their common governance structure is the important feature. Each project remains usable while being shaped through transparent processes that include contributors beyond its originating company.
This arrangement offers a practical alternative to standards being controlled entirely by one vendor. It also acknowledges that enterprises adopting agents need more than model access. They need dependable ways to package organizational knowledge, integrate multiple models, test behavior and limit the consequences of untrusted actions. Shared specifications can reduce integration costs, although they cannot eliminate the work of validating each implementation.
Manik Surtani, the foundation’s chief technology officer and co-founder, described the expansion as evidence that agentic AI needs a neutral home where shared standards, protocols and infrastructure can develop. His point is not that open governance removes competition. It is that the basic interfaces can be developed collectively while companies continue to compete in models, hosting and finished applications.
There are meaningful tests ahead. Developers will determine whether the specifications remain simple enough to implement, whether security guidance arrives quickly enough and whether competing projects can evolve without governance becoming too slow. Neutral control can increase trust, but it can also spread decision-making across more participants. AAIF’s first-year growth shows that major companies are willing to test that model, not yet that it has solved every coordination problem.
Agent Skills Brings Procedural Knowledge Into Focus
The donation of Agent Skills carries special weight because the standard already has a sizable user base. Anthropic developed it after observing that capable language models often still lacked the procedural detail required for real business tasks. The answer was a portable folder format that teaches an agent how to perform a specific job and gives it the tools required for that work.
Under the donation, the foundation receives the Agent Skills name and logo, the associated GitHub organization and the agentskills.io domain. Anthropic also transferred maintainership responsibilities, with David Soria Parra and Mahesh Murag named as lead maintainers. Engineers from Anthropic, Google, OpenAI and other companies are expected to participate.
The proposed roadmap includes versioned releases, a formal proposal process and security guidance for authors and installers of skills. That final item is critical. A reusable instruction package can be useful and dangerous at the same time, particularly when it invokes scripts or carries instructions written by unknown parties. Broad adoption increases the value of shared review, but only if security practices develop at a comparable pace.
From Model Interfaces to Controlled Execution
Open Responses reflects an industry reality that agent products rarely depend on a single model. Teams choose models for cost, latency, reasoning performance and domain fit, sometimes changing them as those factors shift. A common interface can make that flexibility less expensive to maintain.
OpenEnv and OpenSandbox address a related production concern. Reinforcement learning requires environments that expose the same behavior to training systems and agent frameworks. Sandboxing isolates code generated or selected by a model before it reaches production systems. Both concerns become harder when every vendor maintains a private implementation with different assumptions.
Containerization and a familiar interface can reduce friction, but they do not automatically produce safe agents. Teams still need resource controls, network policies, monitoring and clear escalation rules. Open specifications can make those controls easier to audit and replace. They can also create a false sense of uniformity if implementers interpret lifecycle or security requirements differently. Adoption metrics will matter, but so will conformance and documentation quality.
A Neutral Test Bed for Agentic Infrastructure
The foundation’s growth from three projects to 10 in roughly a year shows a rapid shift in how agent infrastructure is being built. Vendor-neutral governance is no longer an abstract proposal. Anthropic, OpenAI, Hugging Face and Alibaba have each placed a distinct technical contribution under shared oversight, alongside protocols and tools donated earlier.
For developers, the immediate benefit is a clearer map of the emerging stack. Skills define how capability is delivered. Open Responses defines how applications interact with models. OpenEnv supports training and evaluation, while OpenSandbox governs controlled execution. The existing projects cover agent communication, routing, gateway functions, documentation conventions and agent deployment.
For enterprises, the promise is lower switching costs and more predictable integration. For independent developers, it is access to specifications and reference implementations without needing a commercial relationship with any single provider. For vendors, it is a way to shape common infrastructure while preserving room to differentiate above it.
The announcement was made during AGNTCon and MCPCon North America on October 22 and 23 at the San Jose McEnery Convention Center. That gathering brought together engineers and organizations working on the practical deployment of agents, with sessions focused on implementation experience and common operational challenges. That is the bet the Agentic AI Foundation is making, and the next year of membership will show whether the donated projects convert into production use.
What happens next will determine whether neutral governance can keep pace with the speed of agent development. The Agentic AI Foundation now hosts the infrastructure for that experiment, and its first-year expansion suggests that the industry wants shared ground beneath a rapidly diversifying market.