My agents redo the same workflows in every tool — what does the open Agent Skills standard actually standardize?

October 30, 2026 · 4 min read

SentX SI

It standardizes the package, not the performance. Under the Agent Skills standard, a skill is a folder containing a SKILL.md file that teaches an agent a specific task or domain, along with any scripts or reference material the task needs. That folder is the unit of exchange: write the procedural knowledge once, in one place, and hand the same folder to any platform that accepts the format. Everything else — how each platform reads it, runs it, or extends it — sits outside the standard.

Source: Donating Agent Skills to the Agentic AI Foundation

The origin story explains why the format looks the way it does. Anthropic developed skills while building Claude Code, which handled general knowledge work well but kept hitting the gap between general competence and the procedural knowledge real business tasks require. The feature launched in October 2025, and within weeks developers were publishing thousands of skills on GitHub and writing loaders for Codex, Cursor and OpenCode — community plumbing that arrived before the standard did. Anthropic then published Agent Skills as an open standard in December 2025, under its own GitHub organization and domain, so any agent could use it. On October 22, 2026, it went further: Anthropic donated the standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, the body it co-founded with Block and OpenAI the previous year and to which it had already donated the Model Context Protocol. The donation covers the Agent Skills name and logo, the GitHub organization and the agentskills.io domain, where the specification now lives. You can use the standard or contribute to it without joining the AAIF, and maintenance will run through individual engineers from Anthropic, Google, OpenAI and other companies, with David Soria Parra and Mahesh Murag named as lead maintainers.

Source: Donating Agent Skills to the Agentic AI Foundation

The scale figures come from the same announcement, so they carry Anthropic's word: 1 million skills published, nearly 280 million installed, and more than 70 agent platforms integrated, including ChatGPT, Cursor, GitHub Copilot and VS Code. Whether those numbers survive independent scrutiny is a question the announcement does not invite you to ask, but they do establish the market condition the standard is aimed at — a large installed base of single-tool workflows whose owners would rather maintain each one once.

Source: Donating Agent Skills to the Agentic AI Foundation

This is where the practical reading has to slow down, because "integrated" is doing more work than the announcement defines. The page does not say what integration technically means on any given platform: whether the loader reads SKILL.md directly, wraps the folder in its own metadata, executes the bundled scripts, or ingests only the reference text. It makes no claim that a skill behaves identically across tools, and it offers no quality guarantee for any individual skill — a gap that matters when the announcement counts 1 million published folders and describes no review process for them. The roadmap Anthropic proposed alongside the donation — versioned releases, a formal proposal process, security guidance for people who write and install skills — is proposed, not adopted. What the standard establishes is a common format and a neutral home; what it does not establish is equivalent behavior, and that difference is exactly the distance your agents currently cross by redoing the same work in every tool.

Source: Donating Agent Skills to the Agentic AI Foundation

For a team running several platforms, the way to convert the standard from a promise into a measurement is a three-step portability check. This is our editorial suggestion rather than a tested protocol, and it is deliberately small. First, pull the specification from agentskills.io and compare it side by side with each platform's own skill documentation. Note where the platform accepts the folder as-is, where it demands extra fields, and where it quietly ignores parts of it. Second, choose one skill you actually rely on — a real workflow with known inputs and expected outputs, not a demo — migrate it to a second platform, and run it end-to-end. Third, record where the outputs diverge: not stylistic differences in phrasing, but structural ones — a reference file the agent never read, a script that did not execute, a step the workflow dropped. Repeat for each tool in your stack, and the divergence log becomes your honest inventory of which workflows genuinely travel and which still need local rework. One clean migration proves nothing about the next one; the check earns its keep only as a per-platform, per-skill measurement.

That discipline mirrors a line we put in our own product documentation. The SentX model card advises keeping a portable checkpoint of checked findings and open questions for work across conversations, and adds that the checkpoint builder formats your notes; it does not verify them. A shared folder standard works the same way. The structure travels cleanly — that is the part the standard owns — and the verification stays with whoever depends on the result. If your goal is to stop rewriting the same workflow for every tool, Agent Skills gives you the most credible common ground available today: a neutral foundation, multi-vendor maintenance, and a format already present in the major coding and chat platforms. It gets you from five copies of a workflow to one folder and five test results. Which of those tests pass is still yours to find out.

Source: Victoria: public model card