The Agentic AI Foundation (AAIF) now hosts four more projects, and four rival companies contributed them. On October 22 at AGNTCon + MCPCon North America in San Jose, the Linux Foundation-hosted group took in Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. The foundation launched with three projects in December 2025 and now has ten. The new additions cover most layers of a modern AI agent stack: what an agent knows, how it calls models, where it is trained and tested, and where it runs code.
Four Open Building Blocks for AI Agents
Each new project covers a different layer that developers have been rebuilding, one vendor at a time.
- Agent Skills (Anthropic) is an open format for packaging instructions, scripts and resources that an agent loads only when a task needs them. Specialised know-how can then move between agent products instead of being tied to one assistant. Unite.AI reports the format is licensed under Apache 2.0 and CC BY 4.0. The foundation says it has more than 70 supporting platforms, over a million published skills and nearly 280 million installations, all self-reported figures.
- Open Responses (OpenAI) is a specification for interoperable, multi-provider LLM interfaces, based on the OpenAI Responses API. It defines a shared request schema and consistent streaming events, so developers can write a call once and send it to different model providers. Databricks Model Serving and Vercel AI Gateway are cited as existing implementations.
- OpenEnv (Hugging Face) is a framework and community hub for reinforcement learning environments. It packages environments as containerised services that expose a standard agent interface, MCP and Gym. Its technical committee includes members from Microsoft, Meta-PyTorch, Modal, Prime Intellect and Mercor, among others.
- OpenSandbox (Alibaba) is a general-purpose sandbox platform for running agent code safely. It includes multi-language SDKs, unified lifecycle and execution APIs, and Docker and Kubernetes runtimes. It targets coding agents, GUI agents, agent evaluation and RL training.
The four join the foundation's existing projects: the Model Context Protocol (MCP), Agent2Agent (A2A), AGENTS.md, goose, agentgateway and Agent Router.
From Three Founding Projects to a Full Stack
The Linux Foundation announced AAIF on December 9, 2025. Its founding contributions were MCP from Anthropic, goose from Block and AGENTS.md from OpenAI. AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI joined as platinum members. In less than a year, the group has gone from looking after a single connectivity protocol to running a broad portfolio of agent standards.
This round is notable because the contributors are direct competitors. OpenAI is placing a specification derived from its own API under neutral governance. Anthropic is giving up control of its skills format. Alibaba also framed its contribution as a governance decision. Xiaobin Xu, its head of development infrastructure, said the company does not want OpenSandbox's future "to be defined by a single company."
AAIF CTO and co-founder Manik Surtani described the foundation's original premise as a bet that agentic AI needed a neutral home. With ten projects in its first year, the foundation can now point to evidence that the bet has paid off.
Maturity Varies Across the New Projects
The four projects are at very different stages. Open Responses now has a charter, an Apache 2.0 license and an acceptance test suite that can be run against any vendor's endpoint. That gives it a practical way to check whether a provider really conforms. Agent Skills is a deliberately short, stable folder format.
OpenSandbox and OpenEnv are younger projects that are still changing quickly. Analysts also caution that neutral governance is not the same as portability. A skill written for one agent may still need changes to run in another, and the same request can behave differently on two models even when both accept the same schema.
Why It Matters
For enterprises building agentic AI systems, the biggest hidden cost has been lock-in at every layer. Prompts and procedures are tied to one assistant, integration code is tied to one model API, and test harnesses and sandboxes are built in-house. AAIF now holds open, vendor-neutral options for each of those layers. That makes it more realistic to switch model providers, reuse agent capabilities across platforms and compare agents in shared environments.
The timing also matters for AI agent security. Recent weeks have brought reports of coding agents hijacked through malicious instruction files and of agents behaving unpredictably in testing. A shared sandbox protocol and standard evaluation environments give the industry common tools for containing and measuring agent behaviour, rather than leaving each team to build its own.
The open question is whether these standards will converge in practice or split into vendor-specific variants. For now, the AAIF portfolio is the clearest sign yet that the major agent platforms see interoperability as worth giving up some control.
