Anthropic and OpenAI move agent specs to neutral AAIF foundation
Anthropic and OpenAI contributed Agent Skills and Open Responses to the Agentic AI Foundation, joining Hugging Face and Alibaba in a push for vendor-neutral standards.
In late October 2026, Anthropic and OpenAI transferred key agent specifications to the Agentic AI Foundation (AAIF), a neutral body backed by the Linux Foundation. This move joins earlier contributions from Hugging Face and Alibaba, signaling a major shift toward open governance in the rapidly evolving agentic AI stack.
What happened
The Agentic AI Foundation was established in December 2025 to provide shared infrastructure for AI agents outside the control of any single vendor. Since its launch, it has grown to host ten projects, with significant additions announced on October 22, 2026. Anthropic contributed Agent Skills, while OpenAI contributed Open Responses. These join existing projects like the Model Context Protocol (MCP), which connects AI applications to external tools and data.
Manik Surtani, co-founder and CTO of AAIF, discussed these developments at AGNTCon + MCPCon in San Jose. He emphasized that while the technologies were already in use, their governance under a neutral foundation changes how developers can trust and shape them. The foundation now includes heavy-hitting projects from four major industry players, aiming to reduce fragmentation and vendor lock-in.
To accommodate the fast pace of innovation, AAIF introduced a new Sandbox lifecycle phase on September 1, 2026. This category allows early-stage, high-momentum community tools to develop under neutral governance before reaching enterprise scale. This adjustment addresses feedback that the initial entry barriers, modeled after the mature Cloud Native Computing Foundation, were too high for the current speed of AI development.
How it works
Agent Skills, contributed by Anthropic, provides an open format for bundling instructions, scripts, and reference materials. Agents can load these bundles when performing specific tasks, allowing developers to write instructions once and use them across different agents that support the format. This standardization reduces the effort required to adapt workflows for different models.
Open Responses, from OpenAI, is an open specification based on the Responses API. It defines how applications send requests to language models and receive results. By establishing a shared format, it minimizes the translation work needed when switching between compatible model providers. This abstraction helps decouple applications from specific vendor implementations.
Hugging Face contributed OpenEnv, a framework for building and sharing reinforcement learning environments where agents practice tasks and receive feedback. Alibaba added OpenSandbox, a platform for running agent-generated code in isolated environments. Together, these projects create a layered stack that handles instruction formatting, API communication, training environments, and secure execution.
Key details
- AAIF launched in December 2025 and now hosts ten projects, including MCP, Agent Skills, and Open Responses.
- Anthropic contributed Agent Skills and OpenAI contributed Open Responses on October 22, 2026.
- Hugging Face contributed OpenEnv for reinforcement learning, and Alibaba contributed OpenSandbox for isolated code execution.
- A new Sandbox lifecycle phase was introduced on September 1, 2026, to support early-stage projects.
- Over 300 member organizations have joined AAIF in its first ten months, including fintechs and non-tech giants.
- Manik Surtani advises developers to design for model independence using abstractions, proxies, and standardized interfaces.
Why it matters
For software engineers, the primary benefit is reduced vendor lock-in. In the early days of cloud computing, building core infrastructure on proprietary technology carried significant risk. The same dynamic applies to agentic AI today. By adopting standards governed neutrally, teams can switch model providers or update components without rewriting entire architectures. This portability is crucial as the agentic stack remains fluid and undefined compared to mature cloud-native ecosystems.
Context management and memory portability remain critical challenges. Moving context from one model or harness to another is difficult, but standardized interfaces help mitigate this friction. Surtani notes that while change is inevitable, it is becoming easier to manage. With agents capable of writing code and adapting workflows, migrating between standards or providers is less expensive and complex than in traditional software development.
Developers are encouraged to push vendors for open standards and transparent specifications. Trust in the underlying technology grows when the community has a say in its evolution. This collaborative approach ensures that the tools built today will remain viable and interoperable tomorrow, regardless of which specific model provider dominates the market.
What you can do
- Adopt abstractions and proxies in your application architecture to decouple from specific model providers.
- Evaluate Agent Skills for bundling instructions and scripts to improve cross-agent compatibility.
- Implement Open Responses to standardize how your application communicates with language models.
- Use OpenSandbox or similar isolated environments to safely execute agent-generated code.
- Push your vendors to support open standards and contribute to neutral foundations like AAIF.
- Stay prepared for rapid changes in the agentic stack by designing flexible, modular systems.



