Ask Who Can Change Your AI Standard

Anthropic says it is donating the Agent Skills name and logo, GitHub organisation and agentskills.io domain to the Agentic AI Foundation, under the Linux Foundation. The company says people can continue using and contributing to the standard without joining the foundation. Engineers from several companies will maintain it as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. ¹

The announcement also proposes a roadmap covering versioned releases, a formal proposal process and security guidance. Those are proposed next steps, not evidence that all three arrangements are already operating. The foundation now lists Agent Skills as a project. Its published governing board and technical committee are separate bodies; neither list alone establishes the rules for accepting a particular project change. ¹²³

This is a transfer of stewardship for a technical standard, not a new law, a certification of participating products or a warranty covering every skill. The opportunity is to build shared infrastructure with a clearer institutional home. The practical test is whether its emerging processes make decisions understandable and participation useful.

A skill packages instructions and, where needed, scripts or reference material in a folder that an AI agent can load for a task. An agent is software that uses AI to take steps towards a goal. A common packaging format can help different tools find and use the same material. It does not establish that every model will interpret the instructions equally well. ⁴

A new Berkeley preprint, D-Skill, explores a related possibility: comparing stronger and weaker agents' recorded attempts, then turning the differences into written guidance for the weaker agent. Its tests cover spreadsheet editing, text-based household tasks and chess puzzles. The authors report improved results, but the method needs a capable analysing model and cannot supply abilities the receiving model fundamentally lacks. It is early research, not a field trial across organisations. ⁵

The Institute's inference is that reusable operating knowledge could become a more important asset. That makes its maintenance consequential. A shared instruction file may preserve a team's method, yet changes to the model, tools or working conditions can still alter the result. Keep the instruction, its accepted version and its test evidence connected.

Ask the project for its current decision process rather than relying on a sponsor list. Identify who can propose changes, who decides, how disagreements are recorded and how users learn about incompatible updates. Where a process is still developing, record that uncertainty and avoid promising customers a stability guarantee you cannot support.

Then decide what your organisation can contribute. A reproducible failure case, a well-defined industry workflow or a carefully checked translation may be more useful than a broad declaration of support. Allocate staff time and agree what information can be shared. Do not publish customer records, confidential procedures or proprietary material merely to demonstrate participation.

Mutual value should be specific. Maintainers receive evidence that helps them improve compatibility or documentation. Your organisation gains a better understanding of changes and a route to raise problems. Neither contribution nor membership should be presented as privileged approval, guaranteed roadmap control or endorsement by the foundation.

The development originates with a US-based vendor and a foundation whose published technical committee includes major technology suppliers. That does not establish exclusion or capture. It does make it worth asking which operating environments, languages and user needs appear in the evidence used to make decisions. A list of prominent companies cannot answer that question. ³

Recent regional convenings point to different priorities. Abuja's Build 2026 programme, dated 21–22 October, focuses on shared compute, infrastructure, investment and local ownership. The RISE AI summit in Santiago, held on 20–21 October, foregrounds responsible and inclusive adoption across education, business and health. These are organiser descriptions of priorities, not proof of funded infrastructure or successful deployments. ⁶⁷

Across Europe and the UK, Asia-Pacific, Africa, the Middle East and Latin America, the useful contribution will depend on local language, connectivity, sector rules and institutional capacity. A common file format does not harmonise privacy or intellectual-property law. Nor does an English-language contribution channel remove the cost of maintaining multilingual tests. These differences should shape participation and procurement, not become a routine claim that one region speaks for another.

Choose one representative, low-risk workflow and document the accepted instructions and tools. Test whether it still meets the same human-defined outcome in a second supported implementation. Record failures and extra review effort, not only whether the file loads. This is an Institute-proposed acceptance exercise, not a benchmark result or permission to move sensitive work.

Before widening use, ask the accountable owner for three things: the current change process, the evidence supporting portability and a practical contribution the organisation can sustain. Where the answers are incomplete, keep the deployment bounded. Open standards can reduce dependence on one supplier; responsible participation helps prevent that benefit from becoming dependence on an unexamined process.