The AAIF brings reusable agent skills, model APIs, reinforcement-learning environments, and sandboxed execution to the open agentic AI ecosystem.
SAN JOSE, Calif. — At AGNTCon, the Agentic AI Foundation (AAIF) announced that it added four projects from Anthropic, OpenAI, Hugging Face, and Alibaba. These are Agent Skills, Open Responses, OpenEnv, and OpenSandbox. All are existing projects moving into AAIF’s open, vendor-neutral governance, rather than newly developed products.
The common thread binding the projects together is open governance. That is, who controls the interfaces and infrastructure on which developers build agentic applications. Or, as AAIF CTO and co-founder Manik Surtani put it during his keynote, “When one company holds the keys to a project, open is decided one release at a time. A protocol owned by one vendor is private property.”
As AAIF Executive Director Mazin Gilbert said in his keynote, the foundation is “the mutual home where the open-agentic AI stack is being built. Our strategy is based on three pillars. First, open protocols to enable the Internet of Agents. The second one is defining the Agentic stack, being open under mutual governance, and the third one is having a global community where every one of you, every company, has the opportunity to use and share open AI agents.”
Specifically, here’s what these new projects offer AI agent developers and users.
Agent Skills Packages Reusable Expertise
Anthropic’s Agent Skills is an open format for packaging instructions, scripts, and supporting resources that agents can load on demand. Their point is to make specialized expertise reusable across agent products and platforms, rather than requiring developers to reinvent the wheel for each application.
That places Agent Skills at a different layer from the underlying model: It supplies task-specific guidance and resources rather than changing model weights. In the announcement, David Soria Parra, Anthropic’s co-creator and lead maintainer of Model Context Protocol (MCP), explained the motivation: “Today, we’re donating Agent Skills to the AAIF to ensure that anyone building with skills can shape what’s next.”
Open Responses Reduces Provider-Specific API Work
OpenAI’s Open Responses addresses the interface between applications and language models. Based on the OpenAI Responses API, the specification defines shared request and output schemas, consistent streaming events, and extensible tooling. This means developers shouldn’t have to repeatedly write provider-specific integrations. Instead, they can describe requests and outputs once and use them across providers with less translation code.
This is important. A common API can reduce integration work, but it does not make the models behind that API identical. Open Responses targets interface interoperability, not uniform model capabilities or results. Its stated purpose is to simplify building applications that work across multiple model providers.
OpenEnv Brings Training Environments Into the Portfolio
Hugging Face’s contribution, OpenEnv, is a framework and community hub for creating, sharing, and scaling reinforcement-learning environments. These environments provide the settings in which agents learn and are evaluated through interaction.
OpenEnv uses a containerized service exposing MCP as an agent-facing interface and Gym as the training interface used by training libraries. For those unfamiliar with it, Gym and its successor, Gymnasium, provide a widely used Python API for defining and interacting with reinforcement-learning environments. OpenEnv’s goal is to make environments interoperable and easier to build, share, and connect to training systems.
OpenSandbox Handles Execution Infrastructure
Finally, Alibaba’s OpenSandbox provides a general-purpose sandbox platform for AI applications. It includes multi-language SDKs, unified sandbox APIs, and Docker and Kubernetes runtimes, supporting workloads such as coding agents, GUI agents, agent evaluation, AI-generated code execution, and reinforcement-learning training.
At its core is a sandbox protocol defining lifecycle and execution APIs. This enables teams to add custom runtimes. Where OpenEnv focuses on environments for training and evaluation, OpenSandbox focuses on provisioning and managing the places where agent tasks run.
The Future of AI Agents Is All About Open Governance and Standards
The four contributions extend AAIF’s reach across more of the agent development and deployment stack. Next, we’ll see if its open governance model translates into practical interoperability across competing companies and their products.
It’s vital that it does. As Laura Tacho, an AWS senior principal technologist, said at Open Source Summit Europe 2026 in Prague, “We know trust is a blocker to innovation, and the way to solve this trust problem is not just a technical solution, but a structural one, and it requires open standards.”
Tacho compared proprietary governance to walls between systems: “We end up not building open systems; we end up building walls.”
Turning to the summit’s host city, she added, “I think those of us who work in open source also really value the building of bridges, not just between pieces of technology, but across communities and working together.”
