OpenAI, Anthropic, and Google Appear Together Rarely: Agentic AI Foundation Established, Kicking Off the Battle for Open-Source Standards of Agents

Just now, the Linux Foundation officially announced the launch of the Agentic AI Foundation (AAIF for short). According to the announcement, AAIF positions itself as a neutral hosting platform for open - source projects related to AI agents. Almost all global tech giants have signed up to become members of this foundation. Three companies, Anthropic, OpenAI, and Block, as co - founding members, will contribute three major open - source projects, which will form the pillars of the foundation in its early stages.

Currently, the member list of the AAIF Foundation includes Amazon Web Services, Anthropic, Block, Cloudflare, Google, Microsoft, OpenAI, Cisco, IBM, Oracle, Salesforce, SAP, Snowflake, Hugging Face, etc. They will join hands for the first time to jointly formulate open standards for AI agents.

 The core standards for the first generation are set, and no single member has exclusive decision - making power

Currently, the AAIF Foundation is built around three major open - source projects: Anthropic's Model Context Protocol (MCP for short), Block's goose project, and OpenAI's AGENTS.md specification. The three will work together to standardize the interaction between AI agents and external tools and promote the unification of cross - system operation capabilities.

In an interview related to the launch ceremony, Jim Zemlin, the executive director of the Linux Foundation, said bluntly, "Our goal is to avoid the emergence of a 'walled garden' - style proprietary technology stack in the future, where tool connections, agent behaviors, and collaborative scheduling are locked by a few platforms. By integrating these projects into the AAIF, we can formulate exclusive interoperability standards, security models, and best practices for AI agents." At the same time, Zemlin said, "AI is entering a new stage of development. Conversational systems are evolving into autonomous agents capable of collaborative work. In just one year, MCP, AGENTS.md, and goose have become core tools for developers to build the next - generation agent technology."

Among these three technologies, MCP has the highest popularity and maturity. Anthropic open - sourced it a year ago. The core goal of this protocol is to connect AI agents with data sources in a standardized way. Anthropic vividly calls it the "USB - C interface in the AI field." With MCP, developers can quickly connect to all servers compatible with the MCP protocol without having to build customized integration solutions for different databases or cloud storage platforms. Vinesh Sukumar, the head of Qualcomm's AI products, revealed, "Many productivity and content - creation tasks can be completed entirely on edge devices. With the MCP protocol, users can connect to multiple cloud service providers to efficiently handle various complex tasks."

Since its release, MCP has been widely used. According to data from the Linux Foundation, more than 10,000 MCP servers have been deployed. Mainstream products such as Claude, Cursor editor, Microsoft Copilot, Gemini, VS Code, and ChatGPT all support this protocol. Google announced at the 2025 I/O Developer Conference that it would add MCP support to its development tools. Since then, several Google products have successively deployed MCP servers to improve the data access efficiency of agents. OpenAI also completed the adaptation and connection just a few months after the release of MCP.

In response, some netizens joked, "They may have realized that the commercialization of this kind of technology is a mess. And since they are the only lab without an open - source model, they have thrown symbolic benefits to the open - source community."

OpenAI contributed the AGENTS.md specification, which was launched in August 2025 and aims to provide project - specific instruction support for AI programming agents. It is reported that this Markdown - format standard has been adopted by more than 60,000 open - source projects. Mainstream development frameworks such as Cursor, Devin, GitHub Copilot, and Gemini CLI have all achieved compatibility. It is worth noting that OpenAI is also an early supporter of the MCP protocol. Nick Cooper, an engineer at OpenAI, said, "A protocol is essentially a shared language that allows different agents and systems to work together without developers having to repeatedly build integration solutions. We need multiple protocols to create value through negotiation and communication. This openness and interoperability determine that the industry will never be monopolized by a single vendor, a single platform, or a single enterprise."

As the parent company of the Square payment platform and Cash App, the fintech company Block will include the open - source AI agent framework Goose in the foundation's contribution projects. This framework was released in early 2025. By integrating language models, extensible tools, and integration capabilities based on the MCP protocol, it provides developers with a structured solution for building agent workflows. Brad Axen, the head of AI technology at the company, said that the success of goose proves that open - source solutions can compete with proprietary agents in large - scale applications. Currently, thousands of engineers use this framework for coding, data analysis, and document writing every week.

Axen revealed that for Block, open - sourcing goose has dual strategic significance. "Releasing it to the community can attract global developers to participate in optimization. We already have a large number of open - source contributors, and every improvement they make will ultimately benefit the company's business." At the same time, donating Goose to the Linux Foundation allows Block to obtain stress - test feedback at the community level and makes it a practical example of the AAIF vision, an agent framework that can connect to shared components such as MCP and AGENTS.md.

It is reported that the funding source for AAIF is a "targeted fund," and enterprises can donate by paying membership fees. However, Zemlin also emphasized that financial investment does not equal control. "The project roadmap is formulated by the technical steering committee, and no single member has the right to unilaterally decide the development direction."

Why are shared standards crucial for agents?

This year, major enterprises have rushed to integrate generative AI into various products and business processes. At the same time, tech giants have continuously claimed that we have entered the era of AI agents. Although the development path of AI agent models is still unclear, the huge investment of enterprises in this field has given rise to a number of benchmark tools.

A report released by UiPath shows that the adoption rate of AI agents by enterprises is rising rapidly. As of mid - 2025, about 65% of organizations have launched pilot or deployment work on agent systems, and nearly 90% of senior executives plan to increase relevant investment throughout 2026. Data shows that multi - agent systems can significantly improve business performance. Compared with traditional processes, the error rate can be reduced by up to 60%, and the execution efficiency can be increased by 40%. However, the report quotes a recent study from the Massachusetts Institute of Technology (MIT) and points out that only 5% of enterprises have obtained substantial financial returns from AI projects. At the same time, UiPath warns that 96% of IT experts and security leaders are worried about the escalating risks brought by AI agents and call on the industry to take countermeasures as soon as possible.

In this context, tech giants seem to have reached a consensus: to promote industry standardization. Even for the most widely supported MCP protocol, there is still great uncertainty in its adaptation methods for basic technologies such as OAuth. Without industry consensus, the agent ecosystem may fall into a "fragmented" dilemma, where each system operates in isolation and is difficult to interconnect, repeating the decentralized pattern of the early Internet before the popularization of open protocols.

The core mission of the AAIF Foundation is to avoid this risk: by integrating protocols such as MCP, frameworks such as goose, and specifications such as AGENTS.md into a neutral platform for overall management, it promotes the compatibility and collaboration of agent development frameworks, cloud service providers, and developer tools. Its goal is clear: to enable the next - generation AI agents to operate on open and interoperable standards, and the management of these standards is independent of any single enterprise.

The Linux Foundation also said that it will promote the development of these key technologies in the name of openness. However, at the current pace, this organization may eventually incorporate a large number of AI tools in their infancy.

Previously, the Linux Foundation had launched several projects aimed at supporting the neutralization and interoperability development of key technologies. For example, the organization established the Cloud Native Computing Foundation (CNCF) in 2015, initially aiming to support Google's open - source Kubernetes container orchestration system. Now, it has integrated dozens of cloud - computing tools. The certification and training services for these tools provide a stable source of funds for the foundation. However, it is worth noting that Kubernetes was already a mature technology when Google widely promoted it. In contrast, in the current AI field, although technologies such as MCP and AGENTS.md are highly popular, whether they can maintain their importance in the long - term development remains to be seen.

What do people think?

A core question in the industry now is: Can AAIF really become a true infrastructure, or will it just become another "brand alliance"?

In response, Zemlin said, "In addition to the adoption rate of the standards, whether the agent products of global manufacturers are developed and implemented with these shared standards will be the key indicator to measure its early success." In Cooper's view, the continuous evolution of the standards is the core of success: "I don't want these protocols to become rigid and stagnant for two years after being incorporated into the foundation. They need to be continuously iterated and absorb new industry feedback."

Another potential and more subtle impact is that even with an open - governance model, a company's technology implementation plan may become the default standard due to its fast release speed or high market share. However, Zemlin believes that this may not be a bad thing. He explained by taking the history of open - source as an example, "Just as Kubernetes won in the container field, this dominant position comes from the advantages of the technology itself, rather than the forced control of the manufacturer."

At a recent open - source summit, Zemlin proposed that although only a few organizations are currently using MCP in production, 2026 will see a real wave of enterprise automation: multi - agent workflows, learning - based orchestration, verification frameworks, and new combinations of deterministic and non - deterministic systems. He emphasized, "The scale of agent artificial intelligence does not depend on the size of the model. The key lies in how you build the solution."

At that time, Zemlin also shared his views on the current development of the AI industry: "I don't think we are in an artificial intelligence bubble, but we may be in an LLM bubble." He said that although open - source models are cheaper and have almost the same functions, closed - source models still account for 95% of the revenue, resulting in an estimated annual expenditure of up to $24.8 billion on proprietary systems.

Zemlin also highlighted the rise of the PARK technology stack: PyTorch, AI, Ray, and Kubernetes. (Ray is an open - source distributed computing framework designed to simplify the scaling of AI and machine learning [ML] workloads.) He believes that just as the LAMP technology stack defined the early Web era, this generation of AI technology stack will also define the future technology stack. He claims that PARK is rapidly becoming the default platform for large - scale AI deployment. He compares this moment to the evolution of the Linux kernel, which was driven by the collective pressure of the global developer community to continuously improve the efficiency of various hardware.

At the same time, some netizens are worried, "Maintaining the protocol may be a thankless task." Some netizens also questioned the current three core standard tools of AAIF. "MCP is outdated, and its efficiency is too low. Currently, MCP is the industry standard, but this may not always be the case." "I tried goose, but it can't do anything."

Many people in the community have begun to expect the AAIF to make some contributions: "Is it possible for the AAIF to formulate a community - shared standard similar to the chat completion JSON API? Since everyone seems to be cloning it, it would be great to have a recognized standard specification and a corresponding consistency test suite."

Anyway, for developers and enterprises, the short - term value of the AAIF Foundation is obvious: to reduce the development time of customized connectors, improve the predictability of agent behaviors across codebases, and simplify the deployment process in high - security environments. Its long - term vision also seems worthy of expectation: if tools such as MCP, AGENTS.md, and Goose become the industry's standard infrastructure, the AI agent field may shift from a closed - platform model to an open, compatible, and freely combinable software ecosystem.

Reference link:

https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation

https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation

https://techcrunch.com/2025/12/09/openai-anthropic-and-block-join-new-linux-foundation-effort-to-standardize-the-ai-agent-era/

https://thenewstack.io/linux-foundation-leader-were-not-in-an-ai-bubble/

This article is from the WeChat official account "AI Frontline". Compiled by Hua Wei. Published by 36Kr with authorization.