The next phase of AI is moving from conversation toward action. As agents begin connecting to tools, data, repositories and workflows, the Linux Foundation is creating a new venue meant to keep that ecosystem from becoming a collection of incompatible products.
The new group is called the Agentic AI Foundation, or AAIF. Its launch brings together donated projects from Anthropic, Block and OpenAI, with other members including AWS, Bloomberg, Cloudflare and Google.
Why AI Agents Need Shared Plumbing
AI agents are different from simple chat interfaces because they are designed to do work across software systems. That makes interoperability more important. If each vendor builds its own closed approach for tools, instructions and orchestration, developers may be forced into repeated custom integrations.
AAIF is meant to provide a neutral home for open source agent projects. The idea is not just to collect software under one brand, but to coordinate the pieces that help agents communicate, follow instructions and connect to the systems they need.
At launch, the foundation is centered on three contributions:
- MCP, Anthropic’s Model Context Protocol, which is described as a standard way to connect models and agents to tools and data.
- Goose, Block’s open source agent framework.
- AGENTS.md, OpenAI’s instruction file that developers can add to a repository to guide AI coding tools.
Together, these projects point to a practical goal: reduce the amount of one-off work required to make AI agents useful across different environments.
The Open Source Strategy Behind AAIF
The Linux Foundation is positioning AAIF as a way to avoid a future dominated by locked-down agent stacks. Jim Zemlin, executive director of the Linux Foundation, described the risk as a world of “closed wall” proprietary systems where tool connections, agent behavior and orchestration are controlled by a small number of platforms.
Zemlin said the group can coordinate “interoperability, safety patterns, and best practices specifically for AI agents.” That framing matters because agent systems do not only need to connect; they also need predictable behavior as they operate across software, data and organizational boundaries.
OpenAI engineer Nick Cooper described protocols as a shared language for agents and systems. In his view, the agent ecosystem will need multiple protocols that can negotiate and work together rather than one company controlling the full stack.
That is the broader bet behind AAIF. If the core pieces remain open and vendor-neutral, developers and enterprises may be able to assemble agent systems from shared components instead of committing everything to one provider.
What Each Donated Project Adds
Anthropic’s MCP contribution sits at the protocol layer. The aim is to make MCP neutral infrastructure for connecting AI models to tools, data and applications without requiring endless custom adapters.
MCP co-creator David Soria Parra said the main goal is enough adoption for MCP to become the de facto standard. He also described the benefit for developers: build an integration once and use it across any client.
Block’s Goose contribution comes from a different angle. Block is known as the fintech company behind Square and Cash App, but Goose gives it a role in agent infrastructure. AI tech lead Brad Axen said thousands of engineers use Goose weekly for coding, data analysis and documentation.
For Block, open sourcing Goose creates a feedback loop. Axen said that putting it in the world gives other people a place to help improve it, and those improvements can return value to the company.
OpenAI’s AGENTS.md is simpler but still important. It gives developers a way to add repository-level instructions for AI coding tools. In an agent-heavy environment, that kind of instruction file can help make behavior more consistent across codebases.
Governance Is The Test
The Linux Foundation already hosts major AI and developer infrastructure projects, including PyTorch, Ray and Kubernetes. AAIF is narrower by design. It is focused on agent standards, orchestration, safety patterns and interoperability.
The group is funded through a “directed fund,” with companies contributing money through membership dues. Zemlin argues that funding does not mean control. According to the source article, project roadmaps are set by technical steering committees, and no single member receives unilateral authority over direction.
That governance model will matter because open standards can still produce winners. One implementation may become the default if it ships quickly or gains wide usage. Zemlin compared that possibility to Kubernetes “winning” the container race, saying dominance can emerge from merit rather than vendor control.
The key question is whether AAIF becomes working infrastructure or just another alliance of familiar company names. Zemlin said an early sign of success would be shared standards being developed and implemented by vendor agents around the world.
What Developers And Enterprises Could Gain
For developers, the short-term appeal is direct. Shared agent standards could mean less time building custom connectors and more confidence that an agent will behave consistently across tools and repositories.
For enterprises, the value is also operational. Security-conscious environments often need predictable deployment patterns. If agent frameworks, instruction files and connection protocols become common infrastructure, adoption may be easier to manage.
Cooper said success should not mean protocols frozen in place. He said he does not want them to sit in the foundation for two years without progress, and that they should keep evolving with additional input.
That ongoing evolution is central to the promise of AAIF. If MCP, AGENTS.md and Goose become widely used open building blocks, the agent market could move toward a mix-and-match software model instead of a set of closed platforms. The result would be an AI agent ecosystem built less around lock-in and more around shared foundations.