The Agentic AI Foundation (AAIF) just took in four more projects, The Register reports. The foundation was founded in December 2025 under the Linux Foundation as a vendor-neutral home for agent tech, and it launched with Model Context Protocol (MCP), goose and AGENTS.md. With the new four, it now stewards 10 projects.
The four new projects
- OpenSandbox (donated by Alibaba): a way to constrain what an agent can do, using client SDKs, tools, protocols, lifecycle hooks, container runtimes and in-sandbox execution components.
- Agent Skills (donated by Anthropic): a format for text descriptions that guide agents through specific tasks. Anthropic said it’s “committed to keeping Agent Skills open, community-driven, and vendor-neutral.”
- OpenEnv (donated by Hugging Face): a containerized service for building reinforcement learning environments that interoperate and scale.
- Open Responses (donated by OpenAI): an open API spec for building LLM interfaces that work across model providers, standardizing request and response formats, streaming events and tool calls.
The growth
AAIF has passed 300 members in under a year, with 60 joining this quarter. Gold members include Ant Group, the Gates Foundation, MongoDB, Nutanix, Royal Bank of Canada, Vercel and Vultr. The Register also notes Axios is reporting that US banks are worried about AI agents interacting with customer accounts without bank authorization, though not necessarily without customer authorization.
My take
I’m happy to see shared blueprints for sandboxing, skills and API formats. Agents that follow common standards are easier to run locally, easier to audit and less likely to lock you into one vendor. It also fits the sandbox question I wrote about last week, and the AGENTS.md rules Mesa just loosened. The banks’ worry is the flip side: the more agents touch real accounts, the more the permission and approval layers matter.
Have a blessed day! And Happy Hacking!
