The infrastructure for artificial intelligence is consolidating around open standards. On October 22, 2026, the Linux Foundation announced that four major technology companies donated core projects to the Agentic AI Foundation. Anthropic contributed Agent Skills. OpenAI contributed Open Responses. Hugging Face contributed OpenEnv. Alibaba contributed OpenSandbox.
This announcement took place at the AGNTCon and MCPCon North America conference in San Jose. It marks a clear shift in how the industry builds tools for agents. The Agentic AI Foundation launched in December 2025 with just three projects. It now hosts ten. The new additions join existing protocols like the Model Context Protocol, Agent2Agent, Agent Router, agentgateway, AGENTS.md, and goose.
The news matters because the agent economy requires shared infrastructure to function at scale. When every provider uses a different format for tools and execution environments, developers waste time writing translation layers. Neutral governance removes that friction. It allows builders to focus on capabilities rather than integration mechanics.
The shift to shared infrastructure
Early agent development was highly fragmented. Builders had to guess which format would become the industry default. They wrote custom integrations for every new model and every new platform. This fragmentation slowed down deployment and made maintenance difficult.
The Agentic AI Foundation exists to solve this exact problem. By bringing competing companies together under a neutral umbrella, the foundation ensures that no single entity controls the underlying protocols. Manik Surtani, the foundation CTO, explained that the goal is "putting the community first by giving it one place to build shared standards" across the industry.
This approach mirrors the early days of cloud computing. Just as Kubernetes became the standard for container orchestration, these new projects aim to become the default layers for the agentic stack. The trade off for companies like Anthropic and OpenAI is a loss of absolute control over their specific formats. The gain is massive ecosystem adoption. When a standard is open, developers trust it enough to build businesses on top of it.
Standardizing how agents learn and act
Two of the new projects focus directly on agent capabilities and training. Agent Skills, originally created by Anthropic, provides an open format for giving agents new domain expertise. It packages instructions, scripts, and resources into a reusable format. Agents can then load these skills on demand.
This portability is crucial. Without a standard format, specialized knowledge remains locked inside specific platforms. A financial analysis skill built for one agent would need a complete rewrite to work on another. Agent Skills makes that know how portable. It allows developers to build a capability once and deploy it across any system that supports the standard.
OpenEnv tackles a different part of the capability stack. Contributed by Hugging Face, it serves as an open framework for creating reinforcement learning environments. It operates as a simple containerized service. It exposes a familiar agent interface through the Model Context Protocol and connects to the Gym training protocol.
Training environments teach agents the capabilities they will use in the real world. OpenEnv makes these environments plug and play. It allows teams to share and scale their training setups easily. The committee behind OpenEnv includes representatives from Microsoft, Meta PyTorch, Modal, and others. This broad backing suggests OpenEnv will quickly become a baseline for agent evaluation and training.
Interoperability and safe execution
The other two projects address how agents communicate and where they run. Open Responses, contributed by OpenAI, is an open specification for building interoperable interfaces across multiple large language models. It is based on the OpenAI Responses API.
Open Responses defines a shared schema, consistent streaming events, and extensible tooling. Developers can describe their requests and expected outputs a single time. They can then run those requests across different model providers with minimal translation. This reduces the brittleness of agent applications. When a new model is released, developers do not need to rewrite their entire prompt and response parsing logic.
OpenSandbox, created at Alibaba, provides the secure execution layer. It is a general purpose sandbox platform for artificial intelligence applications. It includes multi language software development kits, unified sandbox application programming interfaces, and runtimes for Docker and Kubernetes.
Security is the primary concern when agents act on behalf of users. Agents need a safe place to execute code, evaluate data, and run graphical user interfaces. OpenSandbox includes a specific sandbox protocol that defines lifecycle and execution commands. This allows engineering teams to extend the platform with custom runtimes while maintaining a consistent interface.
What changes for builders
If you build tools for agents, the landscape is now much clearer. You no longer have to place bets on proprietary formats. The agentic stack has defined layers for skills, communication, training, and execution.
You can now design your systems around these open protocols. You can package your specialized tools using Agent Skills. You can format your model interactions using Open Responses. You can test your agents in OpenEnv and run them securely in OpenSandbox. This standardization reduces your technical debt and increases your potential user base.
The integration of these protocols creates a predictable flow for agent actions. An application can request a task, the agent can load the necessary skill, and the execution can happen in an isolated environment.
At the protocol level, this standardization means your infrastructure can rely on consistent message formats. Instead of writing custom application programming interface calls for every provider, your systems can exchange standard remote procedure calls.
{
"jsonrpc": "2.0",
"method": "skills/execute",
"params": {
"skill_id": "financial_analysis_v1",
"environment": "opensandbox_kubernetes",
"capabilities": ["read_data", "execute_python"]
},
"id": 1
}This level of consistency allows you to build more complex harnesses. You can connect multiple tools and environments without worrying about the underlying translation mechanics. The focus shifts entirely to the quality of the skills you provide and the logic of your agentic loops.
The foundation of the agent economy
The donation of these four projects to the Agentic AI Foundation is a defining moment for the industry. Anthropic, OpenAI, Hugging Face, and Alibaba have recognized that shared infrastructure benefits everyone.
The agent economy will not run on closed silos. It will run on open protocols that guarantee interoperability, portability, and safe execution. For developers and product teams, the path forward is clear. The tools to build the next generation of agents are now open, neutral, and ready for production scale.