Published: 2026-10-23

Agent Skills joins the Agentic AI Foundation: MCP gets a neutral sibling for know-how and workflows

In short: on 22 October 2026 the Agentic AI Foundation (AAIF), the Linux Foundation fund that governs MCP, announced four new projects. The one that matters most to MCP users: Anthropic is donating Agent Skills, the open format that lets agents load instructions, scripts and reference material on demand. Connecting to systems (MCP) and codified ways of working (skills) now sit under the same vendor-neutral governance.

The announcement came on the first day of AGNTCon + MCPCon North America in San Jose (22–23 October). Alongside Anthropic's Agent Skills, AAIF is taking in Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. That brings the foundation to ten projects, up from three at its launch in December 2025. The same day, AAIF said it now has 305 member organisations, with MongoDB, Nutanix, Royal Bank of Canada, Vercel and Ant Group among the new Gold members.

What are Agent Skills, and how do they relate to MCP?

According to the specification at agentskills.io, a skill is a folder containing a SKILL.md file. At minimum it holds a name and a description, plus instructions for one task. The folder can also include scripts, reference documents and templates. At startup an agent loads only each skill's name and description. Only when a task matches does it read the full instructions and, where needed, run the bundled code. That lets an agent keep many skills on hand without filling its context window.

The difference from MCP is the layer they operate on. An MCP server gives an agent access: to your CRM, your accounting system or your document management. A skill captures expertise and process: how your organisation drafts a quote, reviews a contract or formats a monthly report. In practice they complement each other. A skill can specify which MCP tools an agent should use, and in what order.

What are the other three new projects?

Open Responses (OpenAI): an open specification for a common interface to language models, based on OpenAI's Responses API. The aim is to write a request once and run it against multiple providers.
OpenEnv (Hugging Face): a framework for reinforcement learning training environments. Those environments expose, among other things, an MCP interface.
OpenSandbox (Alibaba): a platform for isolated execution environments for AI applications, with Docker and Kubernetes runtimes and a protocol for the sandbox lifecycle.

AAIF already hosted MCP, Agent2Agent (A2A), AGENTS.md, goose, agentgateway and Agent Router.

Why does this matter for businesses?

Two reasons. First, it reduces dependence on a single vendor. Skills your team writes today now follow an open format with neutral governance, just as MCP has since December 2025. That makes internal investment in them easier to justify. Second, the agent stack is becoming easier to map. Connectivity (MCP), process (skills), execution (sandboxes) and model access (Open Responses) each get their own open standard under one roof.

Not everyone is unreservedly positive about that direction. On 20 October Figma co-founder Dylan Field told TechCrunch that MCP is "not a perfect protocol": in his view it favours the frontier AI labs, and opening up more data through MCP increases the risk of abuse. Figma did announce twenty new MCP partners that same week. Neutral governance does not by itself settle who gets access to which data.

Our advice

(1) Treat skills as code, not documentation. A skill can contain scripts that an agent executes, with the same permissions as the MCP servers it uses. Third-party skills therefore deserve the same review as a new MCP server: provenance, owner, version. (2) Capture your own workflows as skills in a version-controlled repository, with an owner per skill. That makes agent behaviour auditable, which helps with the transparency that the EU AI Act and internal auditors ask for. (3) Read the move to AAIF as a governance signal, not a quality label. An open standard says nothing about the security of an individual skill or server. More in MCP security and MCP and EU regulation.

What does this mean for MCP itself?

Our take: nothing changes in the MCP specification right now. No new version was announced; the current one remains the 28 July 2026 release. But MCP is becoming less of a standalone protocol and more one building block in a wider set of open standards. David Soria Parra, co-creator and lead maintainer of MCP at Anthropic, draws that line himself in the announcement: AAIF gave MCP a neutral home, and now Agent Skills follows. If you have an MCP strategy, this is the moment to decide how you will manage the know-how and workflows around it too.

Sources

Frequently asked questions

What is the difference between an MCP server and an Agent Skill?

An MCP server gives an agent access to a system, such as a CRM or accounting package. A skill is a folder with a SKILL.md file that describes how to carry out a task, optionally with scripts and reference material. A skill can tell an agent which MCP tools to use.

Does this announcement change the MCP specification?

No. AAIF did not announce a new MCP version; the current specification remains the 28 July 2026 release. MCP and Agent Skills are now governed by the same vendor-neutral foundation.

Are third-party skills safe to use?

Not automatically. A skill can contain code that the agent runs. Review third-party skills the way you would review a new MCP server: provenance, owner, version and what permissions the agent holds.

Last updated: 2026-10-23