AI skills start on laptops. Enterprises inherit the mess.

Every company adopting AI agents seems to be running the same experiment right now: someone writes a skill file — a runbook, a checklist, a style guide — drops it in a repo or a Slack thread, and forgets to tell anyone else it exists. Sagar Batchu, founder and CEO of Speakeasy, put it bluntly to The New Stack: it's unrealistic to expect an admin to have context over how every skill in a company should be optimally performing. Multiply that by every developer who's ever had a good idea alone at their laptop, and you get chaos. Speakeasy says one mid-sized fintech customer found more than 500 unique AI skills floating around the organization, with no reliable way to know which version was current or who owned what.

The root cause is almost sociological. AI experimentation has always felt personal first — even Claude Code talks to you like it's just the two of you in a chat window. So people don't rush to publish their prompts for the team to see, and the result is a lot of quiet reinvention: five people solving the same problem five slightly different ways, none of them aware the others exist.

Speakeasy's answer, launched Friday under the name Skills Management, treats a skill the way git treats code. Every skill gets registered centrally, every edit produces a new immutable version, and admins get a shared view of what actually exists across the company. Layer on observability — which skills get used, how often, by which teams — and suddenly you can spot the duplication instead of just suspecting it. If a skill starts underperforming, an LLM judge can review it and suggest fixes. The whole thing enforces the emerging Agent Skills specification, plain Markdown with YAML frontmatter, using the frontmatter name as the single source of truth so duplicate content under a different name just gets ignored.

Versioning is where the design gets genuinely careful. Publish a new version of a skill and the old ones get automatically unpublished, which yanks them off employees' machines and lets the observability stack flag any stragglers — lagging plugin versions, deliberate local copies, that sort of thing. Batchu says there's deliberately no version pinning, for security reasons, though rollbacks are possible where it makes sense. It's a system built around the reality that skills start life on someone's laptop and have to be promoted through the organization to become trustworthy — which also means the tooling has to serve very different audiences, from the security reviewer who should be able to invalidate a skill without editing its contents, to the manager who just wants a dashboard showing how promotion is going.

The part that should reassure skeptics is portability. Because skills follow the open Agent Skills format, Speakeasy stores nothing proprietary, and Batchu says users can export their whole skills corpus or clone everything to a GitHub repo whenever they want to leave. That matters more than it sounds, because enterprises are simultaneously trying to rein in an explosion of internal agents while questioning whether the token costs behind them are paying off. Tools that make that value visible — rather than just adding another layer of AI tooling to manage — are going to matter a lot more than the skills themselves.