Anthropic, OpenAI Donate Agent Skills to Agentic AI Foundation

The open agent ecosystem got its biggest standardization boost yet this week. At AGNTCon + MCPCon North America in San Jose, the Agentic AI Foundation announced four new hosted projects, donated by Anthropic, OpenAI, Hugging Face and Alibaba. The move brings the foundation's portfolio to ten projects in its first year, up from three at launch in December 2025.

The announcement matters because of who donated what. Each of the four new projects covers a different layer of the agentic stack: how agents gain new skills, how applications talk to models across providers, how agents are trained and evaluated, and where agents run safely. By placing them under vendor-neutral governance instead of keeping them inside corporate walls, the donors are betting that shared standards grow the whole market faster than proprietary control.

Agent Skills Tops the List of Four New Projects

The headline project is Agent Skills, contributed by Anthropic. Agent Skills addresses one of the most practical gaps in agent development: the lack of a standard way to package, share and apply reusable capabilities to agents.: the lack of a standard way to package, share and apply reusable capabilities to agents. Today, developers who want to extend an agent with domain expertise tend to invent their own formats and wiring. Agent Skills aims to make capability packages portable across platforms, so a skill built once can be loaded by agents running in very different environments.

That idea resonates with anyone building for the agent economy. If Agent Skills succeeds as a standard, independent developers could publish skills the way mobile developers publish app-store apps, and agent operators could pick from a shared library instead of rebuilding the same integrations. The foundation's leadership framed the donation as evidence that open standards are winning: four rival companies each chose neutral governance for work they could have kept proprietary.

The other three projects fill in the rest of the stack. Open Responses, contributed by OpenAI, defines a common schema for how applications talk to models across providers. As agents increasingly need to call on models from more than one company, a shared response format means developers can switch or combine providers without rewriting their integration code. OpenEnv, from Hugging Face, is an open framework for building scalable reinforcement learning environments for training agents. It is designed to make the training and evaluation of agents reproducible and shareable. OpenSandbox, contributed by Alibaba, is a general-purpose sandbox platform with SDKs for multiple programming languages, giving developers a shared way to execute agent workloads safely.

Why Neutral Governance Matters for Agents

The four projects join an already significant lineup. The foundation's existing projects include the Agent2Agent (A2A) protocol, the Agent Router, agentgateway, AGENTS.md, the goose agent framework, and the Model Context Protocol (MCP). Together they now cover the full lifecycle of an agent: how it is described, how it communicates, how it learns, how it is executed, and how it behaves safely.

According to the foundation's announcement, each project operates under open, vendor-neutral governance, which means no single company controls the outcome. That structure is deliberate. The foundation's chief technology officer described the mission as giving the community one place to build shared standards, protocols and infrastructure, with community interests ahead of any single company's control. Reported coverage of the announcement echoed that framing: this is a bet that agentic AI needs a neutral home to thrive.

The timing is not accidental either. The announcement landed at the conference that gathers the open agentic AI community to share implementation experience and discuss common challenges. Events like the foundation's earlier milestone of ten projects show how quickly the governance model has scaled, and this week's donations accelerate that trajectory with direct contributions from the biggest model labs.

What It Means for Agent Builders

For developers building agents, the practical takeaway is convergence. Standards like Agent Skills and Open Responses reduce the cost of building on top of multiple models and platforms. OpenEnv gives researchers a shared, reproducible way to train agents with reinforcement learning. OpenSandbox offers a common execution environment. Each standard that lands makes the next one easier to adopt, because the ecosystem starts speaking the same language.

There is also a competitive signal. When Anthropic and OpenAI both donate core infrastructure to a neutral foundation, it suggests they see the agent layer as infrastructure that benefits everyone, including themselves. That mirrors earlier waves of technology: the internet's protocols, the web's standards, and cloud-native computing all grew fastest once their foundations were open.

The broader trend is clear across recent coverage of the agent economy. Infrastructure is being built specifically for agents: data networks designed for agent access and new compute models are landing every week. Standards are the layer that ties that infrastructure together, and this week's donations put four more building blocks under neutral governance.

The foundation's portfolio has now tripled in a year. With Agent Skills, Open Responses, OpenEnv and OpenSandbox added, the open agent stack is looking increasingly complete. The next test is adoption: whether agent builders actually build on these standards, including Agent Skills. Given who donated them, that seems like a bet worth watching.

Sources: Agentic AI Foundation announcement, Linux Foundation press release, PR Newswire via Morningstar.