--- title: "Context Needs Its Iceberg Moment" url: "https://atlan.com/context-and-chaos/issue/context-needs-its-iceberg-moment/" description: "Context has open formats. Now it needs the open layer above them, the way data needed Iceberg. Prukalpa Sankar on ai-registry and owning context." keywords: "Context Engineering, AI Agents, Enterprise AI, Open Standards" --- A Context & Chaos PK Editorial by Prukalpa Sankar, Co-founder & Co-CEO of Atlan, adapted from her opening keynote at Context Conference. Published November 5, 2026. ## The opening - She opened Context Conference by saying context is under-hyped. Most of the conversation stays at the level of the word; the questions enterprise teams bring have moved past it. - Four of the eighteen questions collected for the keynote: Who owns the context in my company? How do I version a skill? Will it work with the next model? How do I trace a wrong answer back to its context? - Diagram: "The questions teams are asking about context." - The CIO of a global top-10 pharma company: "I know I will have dozens of platforms and hundreds of agents." A head of data and AI at another pharma company runs four agent platforms that "now all speak slightly different versions of the truth." - Enterprise AI is multi-harness, multi-model, and multiplayer. What one person's assistant learns has to become context the whole team can build on, so context has to be portable. The CIO of a global automaker: "I don't want my business IP locked into any one platform." - Diagram: "Enterprise AI across tools, models, and teams." ## Context already has open formats - Agent Skills packages procedures, AGENTS.md briefs coding agents, Google's Open Knowledge Format covers knowledge, and Apache Ossie covers semantics. - Formats let many tools read the same files. They don't say which files belong together or which version is current. ## What Iceberg did for data - Parquet and ORC gave data open file formats, but a folder of readable files can't say which files make up a table, what its schema is, or which snapshot to read. - Apache Iceberg's table spec answers that with shared metadata, and its REST catalog spec gives every engine one way to read and write tables, on storage the company owns. ## The layer context is missing - Diagram: "What Iceberg added for data, and what context still lacks." - Teams keep skills in a Git repo, then post in Slack when one changes. Two colleagues install a week apart, one edits locally, another tool keeps a months-old copy, and all of them say they have the review skill. - Ben Hudson, who leads product for Atlan's agentic interfaces and open source work, found none of the teams his group talked to had solved distribution. - Two problems with two precedents: npm solved distribution for code, and Iceberg solved interoperability for data. Context needs both. - Diagram: "Two problems, two precedents." ## An open spec for registries - ai-registry (https://github.com/atlanai/ai-registry), a first draft of an open spec under Apache 2.0, announced in the keynote. A registry is files on storage a company already has, with versions, digests, and status. - Skill files are the Parquet files; the spec plays Iceberg's role. A tool builds one integration and can read any registry that follows the spec. - Diagram: "One open format and file layout for every registry." - It covers skills and agents today, with MCP servers and metrics through Apache Ossie planned. ## What a spec can't do - A spec can't approve a procedure or decide whether an exception becomes policy. A version gives a place to start when an agent changes behavior after a model upgrade. - Three practices do the rest: checking how agents use context, controlling changes, and using feedback to improve it. ## Check that agents use the context - SkillsBench 1.1: across 18 model and harness combinations, skill invocation ranged from 46% to 99%. - Liu and colleagues: Claude Opus 4.6 loaded all curated skills in 49% of runs, 31% with distracting skills added, and task pass rates fell from 51.2% to 43.5%. - Vercel: Next.js docs as an on-demand skill scored 53%, the same as no docs; telling the agent to use the skill reached 79%; a compressed index in AGENTS.md reached 100%. - Diagram: "Skill invocation and task performance across the three studies." - Ask three questions of any run: did the agent load the right material, did it follow it, and did it finish the task? Test each artifact where it will run, and rerun when the model or harness changes. ## Change context the way you change code - An artifact needs an identifiable release, a review process, a view of which work a change affects, and a way to pull a bad release. - Diagram: "Ownership, versions, dependencies, testing, and release lifecycle." - Reviewers need to see what an edit means: did a calculation change, did a breakdown disappear, does an exception now apply more widely? Match the machinery to the stakes. ## Route corrections to the owner - A user asks for Q3 net revenue retention, gets 112% from version three of the metric, and replies, "I wanted it by segment." The metric owner decides whether that needs a new version or whether the agent ignored a dimension that already existed. - To tell those cases apart, the owner has to see actors, artifacts, and activity together. - Diagram: "Actors, artifacts, and activity connected." - Owners need patterns tied to their artifacts, not every conversation. A reviewed fix ships as a release, and tools that read the registry pick it up. ## What owning context takes - Registries, tests, reviews, and feedback loops protect how a company does business: its definitions, procedures, exceptions, and judgment about its customers. - Start with one recurring agent task: follow a recent run back to the context it used, check which versions it touched, and ask who would approve a fix and how a teammate on another tool would get it. - "Rent the intelligence. Own the context." ## The keynote deck - The 32-slide Context Conference opening keynote deck, as a PDF: https://atlan.com/docs/context-conference-2026-keynote-prukalpa.pdf - Watch the full keynote: https://atlan.com/context-conference/opening-keynote/ ## Next steps - [Subscribe to Context & Chaos, the newsletter on context engineering](https://atlan.com/context-and-chaos/) - [Book a demo or talk to Atlan's sales team](https://atlan.com/forms/talk-to-sales-contact/)