The Agentic AI Foundation (AAIF) is embracing four open-source projects from OpenAI, Anthropic, Hugging Face, and Alibaba. The new projects aim to simplify developing, integrating, and deploying AI agents across platforms.
This was reported by SiliconANGLE. The Linux Foundation established the AAIF in December 2025. It started with three projects; the total now stands at ten. The latest additions address challenges developers face, specifically interoperability between AI platforms and managing isolated execution environments.
One of the new projects is Open Responses, which OpenAI is making available. It introduces a general specification for APIs used to access LLMs. Currently, developers often have to modify their applications when switching AI providers. This is because each provider uses its own API structure.
Open Responses aims to reduce this dependency by allowing applications to communicate with different models through a consistent interface. The specifications are based on OpenAI’s existing API design, which may make adoption easier for developers who already use OpenAI’s services.
Anthropic standardizes agent capabilities
Anthropic is introducing Agent Skills. This framework packages instructions, scripts, and other resources AI agents need to perform specific tasks.
Instead of forcing developers to configure functionality separately for each application, Agent Skills provides a general structure for distributing and reusing applications. This can simplify collaboration between development teams. It can also improve consistency when agents execute complex workflows.
MCP is being expanded
Anthropic previously transferred the Model Context Protocol (MCP) to the foundation. MCP provides a standardized way to connect AI applications to external tools and data sources. Agent Skills complements this approach by defining how task-specific applications can be organized and shared.
The two remaining additions address the environments in which AI agents are trained, tested, and deployed. Hugging Face is introducing OpenEnv, which makes AI training more portable across different training systems. This matters especially for developers working with agents that learn to interact with software tools and simulated environments.
Alibaba’s OpenSandbox takes a different approach. It provides APIs to create and manage isolated environments where AI agents can execute code and communicate with external applications.
These environments are becoming increasingly important as AI agents evolve from text generation to autonomous action. Sandboxes enable developers to test these functions while restricting access to production environments and other sensitive resources.
Together, the four projects expand the AAIF’s activities beyond MCP and other existing technologies. They also demonstrate the growing collaboration among competing AI providers. This is especially true for the underlying infrastructure needed to build agentic applications.
By placing these technologies under neutral governance, the foundation aims to reduce fragmentation. It also seeks to prevent developers from becoming overly dependent on individual providers’ proprietary implementations.