On October 22, 2026, during AGNTCon + MCPCon North America in San José, the Agentic AI Foundation (AAIF) announced four new projects that bring its catalog to ten in its first year of existence. Anthropic, OpenAI, Hugging Face and Alibaba each contributed one project. The fact that four direct competitors are placing infrastructure under the same neutral roof deserves attention—but also a reading free of excessive enthusiasm.
What was announced exactly
The four new projects are:
- Agent Skills (Anthropic): defines the folder format
SKILL.mdso that an agent can declare its capabilities in a structured way. - Open Responses (OpenAI): a multi-provider request and response specification derived from OpenAI’s Responses API, designed to decouple application code from a single model.
- OpenEnv (Hugging Face): a framework for reinforcement learning environments focused on agent training and evaluation.
- OpenSandbox (Alibaba): a sandbox platform for running agents in isolation.
These join the six projects already hosted: Agent2Agent (A2A), Agent Router, agentgateway, AGENTS.md, goose and the Model Context Protocol (MCP). Membership exceeds 300 members, with 60 additions this quarter, according to the official announcement.
Why the layer separation matters
What’s most relevant is not the number of projects but that the four new ones address distinct technical layers of the agentic stack:
- Declared capabilities (Agent Skills): how an agent exposes what it knows how to do.
- Model interface (Open Responses): how an application talks to different LLMs without rewriting the call logic.
- Training and evaluation (OpenEnv): how an agent’s capability is trained and measured.
- Secure execution (OpenSandbox): where the agent runs with real isolation.
The existence of projects for each of these layers under a single foundation suggests the ecosystem is beginning to mature in a structured way, not merely accumulating frameworks. That is a useful technical signal for those designing agentic architectures today.
What neutral governance doesn’t solve
It’s worth being precise here, because the technical coverage sources themselves warn of something that headlines tend to soften: placing projects under neutral governance changes who decides their evolution, but does not guarantee they will work together without additional effort.
Put directly:
- A skill written in
SKILL.mdformat for a Claude-based agent will not work without changes in an agent built on another framework that has not yet adopted the standard. - A request written against Open Responses will not produce identical behavior across two different models, because the specification derives from OpenAI’s Responses API and cross-platform adoption is still ongoing.
- Project governance is at an early stage. Proposal PRs on GitHub show active, not stable, specifications.
None of this invalidates the initiative; it simply defines the real scope of the work that lies ahead.
Cisco FAITH and agentic security benchmarking
In parallel, Cisco Foundation AI published the agentic extension of its FAITH evaluation framework, along with the VulnLoc benchmark. The approach is technically interesting: each task gives the agent a snapshot of a repository and a description of a vulnerability class. The agent explores the code with terminal tools and reports the files it suspects contain the issue, or explicitly states it found none. VulnLoc can be run locally with Docker or on Kubernetes via Agent Sandbox, with read-only access to the repository.
This matters because it signals that evaluating agentic capabilities in security contexts is already an operational necessity. The question is not whether an LLM answers questions about CVEs correctly, but whether an agent with tools can reason about real code under controlled conditions. These are different things, and the difference matters when choosing or building agents for security tasks.
Practical recommendations with their limits
With the information available today, these are the actions worth evaluating:
- Explore Agent Skills if you already work with agentic pipelines based on compatible frameworks such as Claude Agent SDK or CrewAI. Caveat: the format is under early governance; do not assume specification stability or immediate cross-platform portability.
- Evaluate Open Responses if your application code has a strong dependency on a single model API. Limitation: behavior across models is not guaranteed; explicit testing per provider is required.
- Run VulnLoc to measure the real capability of your security agents in vulnerability localization. Requires Docker or Kubernetes with an isolated environment. Important: it only covers localization within repositories, not exploitation or response.
- Monitor the AAIF public repository on GitHub. Proposal PRs show the real state of governance. It is the most reliable signal for knowing when a specification has reached sufficient maturity for production adoption.
The Agentic AI Foundation is one year old and has grown from three to ten projects with first-tier contributors. That is an indicator that the industry wants to converge on shared standards. What it does not yet have is the adoption track record of, for example, the Model Context Protocol. Keeping a close watch makes sense; adopting it in production without your own validation does not.
Structured reference for readers and AI agents
{
"format": "rafael-fuentes-article/1.0",
"language": "en",
"title": "Agentic AI Foundation reaches 10 projects: Anthropic, OpenAI, Hugging Face and Alibaba under neutral governance",
"summary": "AAIF adds four projects in its first year: Agent Skills, Open Responses, OpenEnv and OpenSandbox cover distinct layers of the agentic stack, but neutral governance is not automatic interoperability.",
"author": "Rafael Fuentes",
"publicationDate": "2026-10-28T20:01:16.033+01:00",
"eventDate": "2026-10-22",
"url": "https://falifuentes.com/rafael-fuentes-2026-10-28-pm-en/",
"translation": {
"language": "es",
"url": "https://falifuentes.com/rafael-fuentes-2026-10-28-pm-es/"
},
"topics": [
"AI",
"Artificial Intelligence",
"Cybersecurity",
"Cybersecurity News",
"English",
"OpenAI"
],
"sources": [
"https://www.linuxfoundation.org/press/agentic-ai-foundation-announces-four-new-projects-agent-skills-open-responses-openenv-and-opensandbox",
"https://www.unite.ai/agentic-ai-foundation-adds-four-projects-to-reach-10-in-first-year/"
]
}