Anthropic Hands Its Agent Skill Format to a Neutral Foundation
Agent Skills grew from an internal fix for Claude Code's blind spots into a format installed nearly 280 million times. Now Anthropic is giving up control of it — and the reasons say a lot about agent-platform trust.
Independent UpShaqo analysis built from fresh, attributed sources. We explain the impact instead of repeating the announcement.
Read for leverage: focus on the workflow change, the customer problem, and the next action—not only the product announcement.
Anthropic built Agent Skills to solve a problem inside its own product. Claude Code was competent at general coding and reasoning tasks but kept stumbling on the procedural details that actual business workflows require — the specific way a company files a deploy ticket, formats a compliance report, or structures a client onboarding doc. General knowledge wasn't the gap. Institutional knowledge was. That gap, and Anthropic's fix for it, is now the subject of a governance handoff that matters more than the original launch did.
As Anthropic announced, the company is donating the Agent Skills standard — the name, logo, GitHub organization, and the agentskills.io domain — to the Agentic AI Foundation (AAIF), a body operating under the Linux Foundation that Anthropic co-founded with Block and OpenAI. This is the second major protocol Anthropic has handed off to neutral governance; the Model Context Protocol went the same route earlier. The pattern is now a strategy, not a one-off.
What a skill actually is
Strip away the branding and a skill is unglamorous: a folder containing a SKILL.md file plus whatever scripts or reference material an agent needs to execute a specific task or operate inside a specific domain. An agent reading the folder learns not just what to do but how your organization specifically wants it done. That's the entire mechanism — no proprietary runtime, no special API, just a documented file structure any agent platform can parse and load.
The simplicity is the point. Anthropic published the format as an open standard after developers started writing their own loaders for Codex, Cursor, and OpenCode within weeks of the original October 2025 release inside Claude Code. Rather than fight the fragmentation, Anthropic standardized it. Since that formal publication in December 2025, more than 70 agent platforms — including ChatGPT, Cursor, GitHub Copilot, and VS Code — have integrated Agent Skills, according to Anthropic. The ecosystem produced 1 million published skills and nearly 280 million installs in under a year.
The orchestration layer nobody's talking about
Here's the part the announcement doesn't spell out, and where founders evaluating this standard should slow down: Agent Skills defines the packaging format, not the runtime behavior. How a given platform discovers which skill to load for a given user request, how it resolves conflicts when two installed skills address overlapping tasks, and how it sandboxes the scripts bundled inside a skill folder are all decisions left to each implementing platform. That's a feature for interoperability and a liability for predictability — a skill that behaves one way in Claude Code may be interpreted slightly differently inside GitHub Copilot's loader, because the standard governs the container, not the execution contract.
(Analysis: this is the same tradeoff MCP made for tool connections, and it's a defensible design choice for a young standard — but it means any team deploying skills across multiple agent platforms should expect to validate behavior per-platform rather than assume portability.)
A single-vendor standard was a trust problem
Before this donation, Agent Skills carried Anthropic's name on the GitHub organization and the domain. For a format used inside ChatGPT, GitHub Copilot, and other products built by Anthropic's direct competitors, that ownership structure was an awkward dependency. A platform team deciding whether to build deep integration with a rival's format has to ask whether that rival might someday change licensing terms, deprecate features that favor its own product, or simply let the standard stagnate in service of a competing in-house alternative.
Moving Agent Skills under AAIF removes that overhang. Engineers from Anthropic, Google, OpenAI, and other companies will maintain the standard as individuals rather than as representatives of their employers, with David Soria Parra and Mahesh Murag named as Lead Maintainers. Anthropic says the roadmap now under community discussion includes versioned releases, a formal proposal process, and — notably — security guidance for people who write and install skills. That last item is an admission that none of that guidance exists in a formalized way yet, even with 280 million installs already in the field.
Where the format breaks under its own success
A format that scaled to a million published skills in under a year has a discoverability and trust problem baked in, independent of governance. Anyone can write a SKILL.md file and bundle a script inside it; nothing in the structure described by Anthropic requires signing, provenance tracking, or sandboxed execution by default. An installed skill is, functionally, a small program an agent will run with whatever permissions the host platform grants it.
That's not a hypothetical risk category — it's the same class of problem Anthropic's own security tooling is built to catch elsewhere in the stack. The company's parallel Claude Security product now continuously scans connected repositories for vulnerabilities rather than relying on point-in-time audits, precisely because code that looked safe yesterday can change today. Skills, as self-contained executable packages installed at scale across dozens of platforms, will likely need an equivalent discipline — continuous vetting rather than a one-time publish-and-forget model. Nothing in Anthropic's announcement confirms that AAIF will build this; the security guidance item on the roadmap is a stated intention, not a shipped feature.
The adoption calculus for platforms and teams
For agent platforms, the donation lowers the political cost of deeper integration. A format governed by a neutral foundation, with maintainers drawn from Google and OpenAI as well as Anthropic, is easier to justify betting a product roadmap on than one controlled by a single lab. For enterprise teams building internal skill libraries — codifying how your company wants expense approvals routed, how support tickets should be triaged, how a specific compliance checklist gets applied — the practical adoption barrier shifts from "should we trust Anthropic's format" to "do we trust our own skill-authoring and review process," since the standard itself no longer carries single-vendor risk.
Where the opportunity sits
The defensible plays here aren't in the standard itself — it's open and free to implement — but in the layers above it. Skill auditing and provenance tooling, enterprise skill libraries curated for specific industries, and certification services that vet third-party skills before internal deployment are all gaps the base standard leaves open, much as package registries for npm or PyPI spawned their own security tooling markets years after the base format matured. Teams building agent workflows today should treat Agent Skills as stable infrastructure worth adopting, but treat the content of any third-party skill folder with the same scrutiny they'd apply to an unreviewed dependency.