The Agentic AI Foundation has added four projects, bringing its portfolio from three founding efforts to 10 within its first year. Alibaba contributed OpenSandbox, OpenAI added the interoperable Open Responses specification, Anthropic donated Agent Skills and Hugging Face contributed the OpenEnv reinforcement-learning environment framework.
The additions move common agent infrastructure into a vendor-neutral organization, but adoption is uneven. Open Responses already has support from AWS Bedrock, Databricks, Ollama, Nvidia and vLLM, while Anthropic and Google continue to operate their own primary model interfaces.
Security and identity dominated the foundation’s October conference. The Model Context Protocol now supports long-running tasks in Claude Code, and its roadmap includes agent messaging, skills, identity and an HTTP-native transport. OpenAI reported that large tool schemas may be trimmed when exposed to ChatGPT, beginning with argument descriptions, and requires published servers to mark whether tools are read-only or destructive.
Chrome’s proposed WebMCP standard would let websites expose declared tools rather than forcing agents to scrape pages. Google reported actions 2.5 to 7.5 times faster in cited benchmarks, while metadata can flag read-only operations and untrusted content. These mechanisms improve interoperability but do not establish trust by themselves. Deployers still need scoped credentials, confirmation for side effects, audit logs and defenses against malicious instructions embedded in data an agent reads.