---
title: "Managing AI Artifacts Across a Company: Lessons From npm and Iceberg"
url: "https://atlan.com/context-and-chaos/issue/managing-ai-artifacts-lessons-from-npm-and-iceberg/"
description: "Skills and agent definitions now hold how a company works. What npm and Iceberg teach about sharing, versioning and retiring them everywhere."
keywords: "AI Agents, Context Engineering, Open Standards, Enterprise AI"
---

A Context & Chaos guest take by Ben Hudson, a product leader and builder at Atlan, where he leads Atlan's agentic interfaces like its MCP server as well as its open source ecosystem. Published November 18, 2026. His argument: AI artifacts like skills and agent definitions now hold how a company works, but there's no reliable way to share them, keep everyone on the same version, or retire them everywhere. npm solved distribution for code and Apache Iceberg solved interoperability for data, and a registry for AI artifacts needs both.

## The opening

- His team built a tool to find copies of the same skill across harnesses like Claude Code, Codex and Cursor, and flag the copies that had drifted apart.
- The copies came from how skills were shared: someone posts a skill in Slack, each person copies it into every tool they use (one copy per harness, since each reads skills from its own folder), updates reach some copies and not others, and nobody notices because they all have the same name.
- When skills diverge, people and agents follow different instructions and get different results. The goal: everyone across the company has the same set of skills across all their harnesses, in sync whenever a skill is added, updated, or removed.

## What we heard from other teams

- Teams that already use skills at work raised three questions: consistency (which artifacts exist, and which version each person and agent runs), upgrades (how everyone gets the new version in every tool), and deprecation (how to retire an artifact everywhere).
- Nobody had a complete answer to any of the three, for their people or for their agents.
- Many teams had built something close to a registry, a central place to see which skills exist. That solves discovery, rather than consistency or lifecycle. A Git repo of skills still relies on a Slack or Teams post and each person applying the change to every harness.
- Getting the right version to everyone and pulling a bad release is distribution; tools reading skills the same way is interoperability. npm solved a version of the first for code, Apache Iceberg a version of the second for data.

## Lessons from npm

- Before npm, developers copied JavaScript files into each project and passed updates around by hand. Nobody could easily tell which version a project ran.
- npm: publish a version once, every install of that version gets the same bytes, and a lockfile records what each project installed.
- AI artifacts today are where JavaScript was before npm: someone building an agent can't immediately tell which version of a code review skill their agents ran yesterday without digging through traces.
- The lesson: distribution has to be easy. Publishing is one command, installing is one command, so a fix reaches everyone who wants it.

## Lessons from Iceberg

- Engines could read Parquet and ORC for years, but a table was a directory of files with no versions and no safe concurrent writes, and each warehouse had its own table metadata.
- Iceberg defined that layer as an open spec. Its REST catalog spec gives every engine a standard interface to find a table's current version, so engines like Spark, DuckDB, and Snowflake work on one copy of the data on storage the company owns.
- Skills today are where data was before Iceberg: Claude Code and Codex read skills from different folders, so a team using both keeps two copies of every skill. Prukalpa covered the data side in her Context Conference keynote, [Context Needs Its Iceberg Moment](https://atlan.com/context-and-chaos/issue/context-needs-its-iceberg-moment/).
- Formats like Agent Skills and AGENTS.md are open and registries exist, including from AWS and Google Cloud, but there's no standard file format for a registry, so clients need a connector for every registry and metadata can't move between them.
- The lesson: interoperability comes from a shared spec that Claude Code, Codex, Cursor and other clients all read, open enough that those clients agree to read it. One vendor's registry, even open-sourced, is one more island.
- Diagram: "npm solved distribution and Iceberg solved interoperability. A registry for AI artifacts needs both."

## What a registry for AI needs

- Diligently retire skills. Code that nothing calls does nothing, but a model picks a skill from its name and description, so a stale skill that's still installed can still be called. npm's deprecation only warns on new installs; for skills, retirement has to reach every harness.
- Pin versions for every agent. Cloud coding agents start in a fresh sandbox and get whatever shared skill version is current, so you can't keep a team on an older version, stop a bad update halfway, or check which version reviewed a pull request. Every agent should install a pinned version and record which one it got.
- A registry can't solve everything: the same skill can behave differently in two harnesses or with two models, and testing that is a different issue.

## Where we are today

- Atlan open-sourced a first version of [ai-registry](https://github.com/atlanai/ai-registry), an open spec for building a registry on storage you already own, like a Git repo, with a file spec for registries and a CLI that publishes artifacts and syncs them across tools like Claude Code and Codex.
- He invites teams running a registry, building a software factory, or building a harness to share how they handle distribution and interoperability.
- Prajakta Damle from Google, on the Context Conference panel before his session: open ecosystems win. With AI artifacts we're close to day zero, so the goal is to build standards early.
- The longer version is in [his Context Conference session](https://atlan.com/context-conference/lightning-talk/). Issues and pull requests on the spec are welcome.

## Next steps

- To get new writing on context engineering, subscribe to Context & Chaos: [Context & Chaos newsletter](https://atlan.com/context-and-chaos/)
- To see Atlan working on your own data stack, book a demo: [Book a demo with Atlan](https://atlan.com/forms/talk-to-sales-contact/)
