What Is the Agentic AI Foundation (AAIF)?

Quick answer
The Agentic AI Foundation (AAIF) is a Linux Foundation project that hosts open standards for AI agents.
Its projects include MCP, AGENTS.md, Agent Skills, A2A, agentgateway, and other agent infrastructure.
It was created so that no single vendor owns the file formats and protocols agents depend on.
For a developer, the practical effect is portability: the same skill folder, rules file, or tool server can work across multiple agents.
Governance changes slowly; the formats you use today still work, and vendor documentation remains the place to check behaviour.

What does the Agentic AI Foundation do?

The AAIF hosts specifications and reference projects that agents depend on, under neutral governance. Vendors donate a project, engineers from several companies maintain it, and changes go through a public process rather than a single product roadmap.

The project list reads like the parts of an agent stack: MCP for tools and data, AGENTS.md for repository instructions, Agent Skills for packaged procedures, A2A for agent-to-agent communication, and gateways for routing and policy. The foundation does not build an agent; it maintains the interfaces agents share.

Why does neutral governance matter?

A format owned by one vendor is a format the owner can change under you. Neutral governance makes the format a commitment: the specification has a version, a proposal process, and multiple implementations that must keep working.

It also changes who can adopt it. A competitor will not standardise on a rival’s proprietary file layout, but will implement a foundation specification. That is how a single tool’s convention becomes an ecosystem default.

What changed with Agent Skills in October 2026?

Agent Skills moved into the foundation, including the specification, the GitHub organisation, and the project domain. The file format did not change: a directory with a SKILL.md, with name and description as the required fields. Skills written before the donation remain valid.

The announced roadmap adds versioned releases, a formal proposal process, and security guidance for people who write and install skills. That last item matters because a portable skill is also a portable instruction to run something, which makes curation and trust the open problem.

What should a developer watch for?

Follow the specification repositories rather than product changelogs when you need the authoritative format. Vendor docs describe their implementation, including extensions and defaults, which can differ from the spec.

Expect slow, compatibility-oriented changes, and expect security guidance to arrive after adoption rather than before it. Until registries grade skills, the reviewer is the only trust boundary.

Where this bites vibecoders
A generated answer about agent configuration will often name a vendor as the owner of a format it no longer owns. If you are writing documentation for your team, link the specification rather than a product page, or the instructions will silently drift when the implementation changes.

Where AI coding assistants get this wrong

Describing MCP or Agent Skills as proprietary to the vendor that donated them.
Treating a foundation specification as a guarantee of security.
Assuming a project under neutral governance changes quickly.
Citing a product changelog as the authoritative format specification.
Conflating the foundation’s projects with one product’s feature set.
Ignoring extensions a vendor adds on top of the specification.

Checklist

Link the specification, not the vendor docs, for format questions.
Check which projects your agent actually implements.
Read the security guidance when it lands, not after an incident.
Pin skill and rules files in version control.
Re-test behaviour after a major specification release.
Keep vendor-specific extensions documented separately.

FAQ

Does the AAIF own MCP?

MCP is hosted by the foundation, having been donated earlier, and Agent Skills joined it in October 2026. The foundation provides neutral governance; the maintainers still come largely from the companies that built the projects.

Is the foundation a standards body like W3C?

It is a Linux Foundation project, which is a lighter model: projects keep their own maintainers and release process while sharing neutral governance and legal structure. It is not an accredited standards organisation, and conformance testing is not universal.

Do I need to change anything because of the donation?

No. Existing SKILL.md files, AGENTS.md files, and MCP servers remain valid. The change is about who governs future revisions. Teams should pin versions and review changes when they adopt new releases.

Why does one foundation host so many agent projects?

Because they are used together. An agent reads a repository’s AGENTS.md, loads a skill, and calls tools through MCP. Hosting them together keeps the interfaces compatible and prevents any single vendor from controlling the whole stack.