AI's biggest rivals just donated their tools to the commons. At AGNTCon + MCPCon North America in San Jose, the Agentic AI Foundation (AAIF) announced four new open projects — and each one came from a different giant: OpenAI, Anthropic, Hugging Face, and Alibaba.
The AAIF went from three projects to ten in its first year, and this announcement covers four layers of the agent stack at once: how agents gain new skills, how apps talk to models across providers, how agents are trained and evaluated, and where they run safely. In other words, the industry is standardizing the plumbing of AI agents — and doing it in the open.
What got donated
Each company handed over a real, working piece of infrastructure. Agent Skills, contributed by Anthropic, is an open format for packaging agent capabilities — instructions, scripts, and resources an agent can load on demand, portable across agent products. It's the same spec Anthropic launched about a year ago. Open Responses, from OpenAI, is an open specification for building multi-provider LLM interfaces based on the OpenAI Responses API: one shared schema, consistent streaming events, and extensible tooling, so developers can write requests once and run them across model providers.
OpenEnv, contributed by Hugging Face, is an open framework and community hub for reinforcement-learning environments — containerized, plug-and-play services built on the familiar Gym interface and the Model Context Protocol (MCP). And OpenSandbox, from Alibaba, is a general-purpose sandbox platform for AI apps: multi-language SDKs, unified sandbox APIs, and Docker/Kubernetes runtimes for coding agents, GUI agents, agent evaluation, and RL training.
Why this matters
Today, building an agent mostly means locking into one vendor's SDK. These donations push the other direction: vendor-neutral governance where Agent Skills, Open Responses, OpenEnv, and OpenSandbox sit alongside MCP, Agent2Agent (A2A), goose, and AGENTS.md under one neutral roof. "Four different companies chose to contribute them to neutral governance," said AAIF CTO and co-founder Manik Surtani. "No single company controlling the outcome."
This is also the layer the agent wars are quietly being fought at. As proactive agents like Claude Duet and OpenAI's Dots go mainstream, the question isn't just which agent is smartest — it's which infrastructure they all share. Standard sandboxes and shared evaluation environments also matter for safety: you can't build a kill switch for agents without a common definition of where agents run.
AiPost's take
Rivals donating infrastructure is rare — and a signal. When OpenAI and Anthropic both hand their tooling to a neutral foundation, it means the agent race has moved past "who has the best model" to "whose plumbing everyone builds on." For developers, that's good news: skills that work everywhere, APIs that don't lock you in, sandboxes that are actually standard. Watch what gets built on these specs — that's where the next wave of agents will come from.
Comments
Post a Comment