AAIF expands its open foundation for AI agents

The Agentic AI Foundation welcomes four projects from Anthropic, OpenAI, Hugging Face, and Alibaba. The goal is simple: to make AI agents more interoperable, but under open governance.

The Agentic AI Foundation (AAIF) announced at AGNTCon in San Jose the arrival of four projects from Anthropic, OpenAI, Hugging Face, and Alibaba into its open governance. These components cover reusable skills, model interfaces, training environments, and sandbox execution. At this stage, the challenge is not to create new products but to determine who controls the rules of the game for AI agents.

Quatre briques, quatre couches du même empilement

The contributions have different names but aim for the same goal: to make AI agents easier to move from one provider to another. Anthropic brings Agent Skills, an open format to package instructions, scripts, and associated resources. OpenAI adds Open Responses, a specification that harmonizes requests, outputs, and streaming events around its Responses API. Hugging Face pushes OpenEnv, a framework to create and share reinforcement learning environments, while Alibaba offers OpenSandbox, a secure execution platform for agent workloads.

In short, each project occupies a different layer of the agentic stack: skill, interface, training environment, and execution location. This separation matters because it allows components to circulate without forcing developers to rewrite their entire infrastructure each time they change providers. The promise is appealing. The devil, as often, is in the integration details.

Pourquoi l’ouverture compte autant que la technique

The core message of the AAIF is open governance. Manik Surtani, CTO and co-founder of the organization, summarized the challenge as follows: “When one company holds the keys to a project, open is decided one release at a time. A protocol owned by one vendor is private property.” Mazin Gilbert, executive director of the foundation, described AAIF as “the mutual home where the open-agentic AI stack is being built”. The idea is to build a common base rather than a series of gateways dependent on a single player.

This logic mainly targets interoperability. Open Responses, for example, reduces integration work between models but does not make models equivalent. OpenEnv standardizes training and evaluation environments without homogenizing results. OpenSandbox governs where tasks run, with Docker and Kubernetes runtimes, multilingual SDKs, and unified APIs. We are therefore talking about architecture, not magic. Agents will not become wise by decree.

La sécurité des agents passe de plus en plus par la gouvernance

The issue is not theoretical. Companies are already seeking to control what their agents can see, launch, or modify. ServiceNow has unveiled AI Passport, a framework designed to validate authorized accesses and executions for an agent, relying on NVIDIA OpenShell and its own Super Intelligence Control Tower. The goal is to apply business rules at machine speed, with control able to block or authorize an action in less than a second according to the described architecture.

This concern aligns with OpenAI’s tests with independent evaluators. According to METR and Redwood Research, a three-week pilot conducted from July 21 to August 11 gave two researchers on-site access to OpenAI’s code, documentation, and some internal staff. The test identified several new vulnerabilities, including an exploit capable of bypassing monitoring of hundreds of recent actions. OpenAI says it has fixed the identified flaws. Again, the question is not only whether the system works but who can audit it, how, and with what independence.

Ce que cela change pour les éditeurs et les DSI

The rise of AI agents shifts the debate from performance to control. For publishers, standardization reduces porting costs and can accelerate adoption. For IT directors, it enables finer governance, especially when agents access internal tools, sensitive data, or production systems. For regulators, it also offers a clearer target: standards and control points rather than a multitude of closed solutions.

One simple test remains: will these open components really be compatible between competing companies, or only “open” on paper? AAIF bets on a positive answer. Upcoming deployments will show whether this coalition manages to avoid the most classic technical trap: announcing interoperability before resolving the sticking details.

Key points

  • AAIF announced four projects at AGNTCon in San Jose.
  • Anthropic, OpenAI, Hugging Face, and Alibaba are involved.
  • OpenAI conducted an external pilot from July 21 to August 11.
  • ServiceNow talks about governing agents “at machine speed.”
  • Two-thirds of the 3,766 ServiceNow respondents lack automation.
  • The debate focuses on interoperability and open governance.

By the numbers

  • 4 projects — integrated by AAIF at the AGNTCon announcement.
  • 3 weeks — duration of the external pilot conducted at OpenAI, from July 21 to August 11.
  • 61 pages — length of the report delivered to OpenAI management by METR and Redwood Research.
  • 66% — share of the 3,766 security managers surveyed by ServiceNow who say they cannot act automatically.

To read