Context Conference 2026 ran online on October 28, 2026, with eight sessions and nearly 1,500 people watching live. Atlan hosted it, and speakers from Verizon, BNY, Workday, Snowflake, Google, Microsoft and Anthropic worked through the question your team is probably asking now: how do you run context for AI in production?
Together, the sessions covered four parts of that question: who owns context, how teams share it, how it reaches the right agent, and how it moves between tools without lock-in. Atlan, which builds an enterprise context layer, used the day to announce three releases, Atlan Trident, the open AI Registry spec and the open-source Forward Deployed Playbook, plus an accelerator program. Every session is available to watch on demand.
- Enterprise AI is multi-harness, multi-model and multiplayer, so context cannot live inside one tool.
- Verizon, Workday, Upwork and DocuSign described what running context looks like inside their companies.
- Platform speakers backed open formats and split on where governed definitions should live.
- Atlan released a workspace for agents and skills, an open registry spec, and a method for building analytics agents.
| Question | Answer |
|---|---|
| Date and format | October 28, 2026, online |
| Sessions | 8, from a 7-minute recap to a 37-minute panel |
| Live audience | Nearly 1,500, per Austin Kronz’s closing recap |
| Speakers from | Atlan, BNY, UKG, Verizon, Workday, DocuSign, Upwork, Snowflake, Google, Microsoft, Anthropic, plus AstraZeneca’s former CIO |
| Releases | Atlan Trident (private preview), AI Registry spec, Forward Deployed Playbook |
| Program | Analytics Agent Accelerator, invite-only, from November 9 |
| Recordings | On demand at atlan.com/context-conference |
What were the key takeaways from Context Conference 2026?
Five ideas ran through Context Conference 2026, from Atlan’s opening keynote to the closing debate among Snowflake, Google, Microsoft and Anthropic speakers.
- Context is a production dependency. Atlan’s keynote argued it needs owners, versions and releases, as code does; Verizon’s architect described the same need from inside a buyer.
- No single tool will hold it. Verizon runs several agent studios, and Anthropic’s Amandeep Khurana expects context to stay spread across the enterprise.
- Agents read differently from people. In Atlan Labs testing, a table described with its DDL and known traps made an agent about 20 points more accurate than a plain-language description.
- For analytics, context moved accuracy more than model choice. Atlan’s own testing reached 83% with context mined from query history and BI layers, with no human help.
- Open, but not settled. Outside speakers backed open formats and split on whose engine should hold the governed definitions.
| Session | Speakers | The idea to take away |
|---|---|---|
| Enterprise AI is missing a context layer | Prukalpa Sankar | Manage context like code |
| Why invest in context? | Cindy Hoots, Leigh-Ann Russell, Prakash Kota | One word, several meanings |
| Inside Verizon’s context architecture | Steven Huang | One platform, many agent studios |
| The company brain for shared skills and ontologies | Rohan Goel, Ashwin Murali | From single-player to multiplayer |
| The Post-AI Data Stack | Zach Michel and a panel | Metrics, ontology and skills |
| Open sourcing a spec for AI registries | Ben Hudson | A registry is files you own |
| Will open context interchange ever be real? | Bob Muglia and four panelists | A standard with dialects |
| Recap of the day | Austin Kronz | The day in seven minutes |
Why did the opening keynote call context underhyped?
Prukalpa Sankar, co-founder of Atlan, opened Context Conference by arguing that context is underhyped, and spent 21 minutes making the case: “Everybody says today that context is the hottest topic in AI. I actually think it’s underhyped.”
Her evidence started with what holds AI back. In a study Atlan commissioned from MIT Technology Review Insights (2026), 68% of organizations ranked governance, business understanding or trustworthy outputs as their biggest AI barrier, in a survey of 250 executives.
She framed enterprise AI as multi-harness, multi-model and multiplayer, and called the last one the most neglected: “Multiplayer is the most underrated aspect of enterprise AI. Today, even the best AI tools that we have, they’re built for one person.”
Her answer was a nervous system that connects three things: actors (people and agents), artifacts (skills, metrics, ontologies) and activity (traces and feedback). She argued that artifacts need the discipline software already has, a “GitHub for context”, with a lifecycle for context and tests before release.
She also cited a Vercel study on format. In Vercel’s agent evals (2026), the agent never invoked a documentation skill in 56% of cases, while the same docs compressed into an AGENTS.md index reached a 100% pass rate. Her point was that context engineering for AI agents means building context in the form the harness will actually load.
She closed on ownership.
“That is your context. It’s how work gets done. Context is IP. So rent the intelligence but own the context.”
Why are enterprise leaders investing in context?
The executive panel, titled “Why invest in context?”, put that question to technology leaders from BNY and UKG, moderated by AstraZeneca’s former CIO. The conversation stayed on meaning, accountability and the freedom to change tools.
Cindy Hoots, former Chief Digital Officer and CIO at AstraZeneca, opened with a story about her first weeks there: “And then after about 20 minutes, I realized that API to her was not an application programming interface, it was an active pharmaceutical ingredient.” The same word meant two different things inside one company.
Leigh-Ann Russell, CIO and Global Head of Engineering at BNY, argued that people stay accountable for what AI does with data: “human over the loop becomes really important, and we as humans still have to be accountable for our own outcomes of the data.” She warned that “Coherence in AI is really important, because if you have chaos- AI is just going to amplify that chaos.”
Prakash Kota, Chief Transformation Officer at UKG, judged every bet on whether it can be undone: “So I always have this, is it reversible or irreversible kind of decisions. And when you bring in any tech stack, you need to have portability and flexibility in place”. UKG is “rewiring some of the business processes, not trying to put some AI on top of existing process.”
Shared meaning and accountability are governance questions before they are technology ones. Someone has to own the disputed definition, which is the real argument behind who owns AI governance, and someone has to answer for what an agent does with it, which is what an AI governance operating model has to settle and where the return on agent governance comes from.
How does Verizon run one context platform across many agent studios?
Steven Huang, Lead Enterprise Architect at Verizon, showed the Enterprise Context Platform (ECP) Verizon is building so context is not trapped inside each vendor’s agent studio. Verizon runs three kinds of agents: operational, knowledge and analytical.
He described the problem plainly: “If we allow this condition to continue, we’re gonna have problems. Knowledge silos, no sharing, reuse of knowledge, no standards, and we’re vendor locked in. If we decide to move away from any of one of these tools, we will lose lots of context objects we have curated for those projects residing in those tools.”
The ECP is Verizon’s own platform, built on a context object store, a knowledge graph store and curation pipelines, and it runs two of those pipelines:
- Artifacts and tools: standard operating procedures, knowledge articles, MCP tools and APIs, for operational and knowledge agents.
- Assets: metadata, semantic models, reference data and query logs, for analytical agents.
He shared plans rather than results: “Our initial plan is to onboard sixteen thousand business documents and twenty-three thousand knowledge articles”. The ECP serves context and metadata only; it does not proxy database or application calls. Agent builders “come to the ECP, treating ECP as a single one-stop shop, get all the context they need to fulfill their business use case when they develop that agent.”
He named one gap the industry has not closed: there is no standard yet for authoring agent operating procedures, and every agent studio has its own authoring environment. The design matters for its shape: one context layer for many agents instead of one per vendor stack, the trade-off between single-stack lock-in and a neutral layer that sits at the center of agent context layer design.
What does a shared company brain for skills and ontologies look like?
Rohan Goel, Director of Product at Atlan, and Ashwin Murali, Lead Solutions Consulting at Atlan, showed how skills, metrics and ontologies move from one person’s laptop to a whole company. The session used Atlan Trident, which is in private preview.
Many teams already share skills through a Git repository. In Trident, skills arrive from the desktop app, a GitHub repository, the CLI or the API, and each publish adds a version. What it adds around the files is the record: which runs used which version, eval results, and access to models and tools under the caller’s own permissions.
That record showed up in Atlan’s own numbers: “It’s very easy to create new skills. If you see, we have about eleven hundred skills, but only three hundred of them have been used in the last thirty days.”
A quarter-close skill may run only four times a year and still matter, so each unused skill needs an owner who decides whether it stays. That decision is the job of an agent skills registry and of skill best practices.
Ashwin focused on where agents fail: “And what we keep seeing is that this last step is where most agents break, usually because the agent doesn’t really know how your business works.” His answer was an ontology, the business-language core of a knowledge graph for AI agents, which “sits one level above those systems and writes that down in the words that business actually uses.”
His demo, an illustration with a fictional mobile carrier, showed why every rule change has to reach every agent at once: “The business keeps changing. And all these eight agents, your actors, need the rules as soon as they get approved, which is why you need to manage all of these context artifacts like code”. Versioned, governed context for agent skills, kept in a context repository, is how that happens without eight manual updates.
What do analytics agents need to get the numbers right?
Zach Michel, who leads AI solution architecture and support at Atlan, opened The Post-AI Data Stack with what talk-to-data agents need, then a panel from Workday, Upwork and DocuSign took it into practice. The answer was three kinds of context, built and tested on purpose.
He started with a failure most data teams will recognize: “One who’d built an agent and got it to about 89% accurate. That agent, however, still fizzled out and did not gain traction because that specific use case, from a data integrity point of view, needed 98% plus accuracy.” The lesson is to set the accuracy bar per use case before building.
Over 18 months and 12 to 14 iterations, his team found that context mined from query history and BI layers reached 83% with no human intervention, and that “an open-weight model with the right context can be as or more accurate than a frontier model.” He did not name the models or the question set, so treat it as Atlan’s own result rather than a published benchmark.
The context comes in three parts: metrics carry grain, filters, dimensions and verified SQL; an ontology answers “why” questions; skills answer “how” questions. The method runs bootstrap, manage, then improve against evals, and Atlan published it as the Forward Deployed Playbook.
Joe DosSantos, VP of Enterprise Data and Analytics at Workday, recalled that “when I started at Workday, we had a thousand Tableau dashboards, and each one of them was adamant that it was in fact the source of truth.” Sowmya Parthan, Sr. Director of Data Platform and AI Engineering at Upwork, said Upwork built an ontology, domain data and metrics, and now sees accuracy from 70% to 96% depending on the domain, with three domains at 96%.
Piyush Bhargava, Sr. Director of Global Data and Analytics at DocuSign, argued that context lives beyond the warehouse, in support cases and Confluence pages too: “data is a continuum now.” Austin Kronz, Field CDAO at Atlan, summed up the fallback the panel described:
| Route | When the agent uses it | What it tells the user |
|---|---|---|
| Certified metric | A governed definition covers the question | Which certified metric it used |
| AI-ready data | No certified metric covers it, so it browses curated data | That the answer is its best guess |
For questions that need perfect answers, he added, “if you need a hundred percent accuracy on something, dare I say, you might need a dashboard or like take the LLM out of the equation”. That routing is why text-to-SQL for the enterprise depends on a semantic layer built for agents and on evals that measure accuracy per question type, and why the context analytics teams already own is the place to start.
Will open context interchange ever be real?
Bob Muglia, former CEO of Snowflake, moderated a debate with speakers from Snowflake, Google, Microsoft and Anthropic on whether context can move freely between platforms. They agreed on openness in principle and split on where governed definitions should live.
He started with his own history: “So my head of finance and I, we created what we called the Snowflake Bible, which ultimately became about a 60-page document that described all of the core business practices we had”.
Amandeep Khurana, Member of Technical Staff at Anthropic, said: “I don’t think context is going to live inside a single platform, right? It is spread across the enterprise.” That makes context portability a requirement.
On format, nobody defended files alone. Prajakta Damle, Senior Director of Product at Google: “you don’t want the agents to read a revenue definition from a set of markdown files. Like, you want that to be pretty authoritative, governed in a semantic model.”
Dipti Borkar, VP of Microsoft IQ and OneLake at Microsoft, called markdown files “necessary but may not be sufficient”. Khurana had moved his own context layer off files: “So I moved it into Postgres, and all of that went away, but then I had to give up on all the flexibility of files, right?”
The split was over whose engine computes the governed version, the same line that separates context engineering from a semantic layer.
Josh Klahr, Head of Product Management for AI-Powered BI at Snowflake, put it in his engine’s terms: “Here’s a metric name. I’m gonna calculate it with a compilation engine.” Khurana expected dialects: “SQL is a standard, you just mentioned, but every database still speaks its own dialect, right?”
On standards, Apache Ossie (incubating, formerly Open Semantic Interchange) covers governed semantic models and Google’s Open Knowledge Format covers unstructured context. Borkar put interop coverage at “maybe we’re at 20% today”. As with the agent interoperability protocols before them, the number of vendors that implement these specs will decide how far context moves.
What did Atlan launch at Context Conference 2026?
Atlan made three releases at Context Conference 2026 and opened one program: a workspace to manage context, a spec to move it, a method to build it, and a cohort to apply it.
| Release | What it is | Status | Where to find it |
|---|---|---|---|
| Atlan Trident | Where an organization keeps its AI agents and the skills they use, with versions, traces, evals and governed access to models and tools | Private preview | Docs |
| AI Registry | An open spec for agent registries on storage you own, under the Apache 2.0 license | Open source, pre-1.0 | GitHub |
| Forward Deployed Playbook | Atlan’s method for building analytics agents | Open source | analytics-playbook.atlan.com |
| Analytics Agent Accelerator (program) | A cohort led by Atlan’s forward-deployed engineers | Invite-only, from November 9 | Apply |
Ben Hudson, who leads product for Atlan’s agentic interfaces and open-source work, introduced AI Registry with a lesson from software: “NPM solved this in the world of software and made distribution extremely easy. You publish once, everyone installs the same version, you can easily pull bad releases.”
It runs on storage you already own: “a registry is just files on storage, so you can use any storage that you already have, like a Git repo, just as long as the files adhere to the spec.” The spec adds what raw SKILL.md files lack: versions, digests, owners and status, including whether an artifact has been retired. The talk did not cover signing, which is still the open question in whether agent skills are safe for enterprise use.
It supports skills and agents today, and Hudson said MCP servers and metrics, through Apache Ossie, come next. That scope sits between an MCP registry and a model registry.
The MIT Technology Review Insights study and AI Context Platforms For Dummies came out with the event, and the resource pack collects the main resources. Next is the Context in Practice Summit on December 3.
Real stories from real customers: Shared language for AI
Two Atlan customers on the problem the executive panel opened with, from earlier customer stories rather than the conference stage:
"Atlan captures Workday's shared language to be leveraged by AI via its MCP server. As part of Atlan's AI labs, we're co-building the semantic layer that AI needs."
— Joe DosSantos, VP, Enterprise Data and Analytics, Workday
"Context is the differentiator. Atlan gave our teams the shared vocabulary and lineage to move from reactive data management to proactive AI enablement."
— Kiran Panja, Managing Director, Cloud and Data Engineering, CME Group
What running context like software looks like
Two sessions at Context Conference 2026 made the case for context; the other six covered how to run it. Speakers from very different companies described similar practices in different words.
Verizon is building its platform to curate context once and serve it across its agent studios. Upwork measures accuracy domain by domain and reached 96% on three of them. Atlan’s own skills count, 300 of about 1,100 used in a month, shows why every artifact needs an owner.
The pattern is the one Prukalpa Sankar described in the keynote: context gets owners, versions and releases, and it stays yours when the tools change. The Context Conference sessions show what that looks like at Verizon, Workday and Atlan itself, and every one is on demand.
For more events on the same theme, see the context engineering conferences and events of 2026. Which engine holds the governed definition, and how far specs like Apache Ossie get past 20% coverage, is still open.
FAQs about Context Conference 2026
1. When was Context Conference 2026?
Context Conference 2026 took place online on October 28, 2026. Atlan hosted eight sessions over about three hours, covering the opening keynote, an executive panel, Verizon’s architecture, a demo of Atlan Trident, analytics agents, the AI Registry spec, an open-standards debate and a closing recap. Nearly 1,500 people watched live.
2. Can I watch the Context Conference 2026 sessions on demand?
Yes. All eight sessions are available on demand on the Context Conference hub on atlan.com, each with a full transcript. Session pages also link related resources, such as the MIT Technology Review Insights report and the Forward Deployed Playbook for analytics agents.
3. Who spoke at Context Conference 2026?
Speakers came from Atlan, BNY, UKG, Verizon, Workday, Upwork, DocuSign, Snowflake, Google, Microsoft and Anthropic, along with AstraZeneca’s former Chief Digital Officer and CIO, Cindy Hoots. Bob Muglia, former CEO of Snowflake, moderated the open-standards debate.
4. What is Atlan Trident?
Atlan Trident is where an organization keeps its AI agents and the skills they use. It records versions, links runs to traces, runs evals, and gives agents governed access to models, tools, metrics and ontology context. It works beside Claude Code, Codex, Cursor and custom code, and it is in private preview.
5. What is AI Registry?
AI Registry is an open specification for agent registries, released by Atlan under the Apache 2.0 license. A registry is a set of files on storage you already own, such as a Git repository, with versions, digests, owners and status added to each skill or agent. It supports skills and agents today.
6. What is the next Atlan context event?
The next event is the Context in Practice Summit on December 3, 2026, where customers share how they run context for AI agents in production. Austin Kronz announced it in the closing recap of Context Conference 2026, and registration is open on atlan.com.
Sources
- AGENTS.md outperforms skills in our agent evals, Vercel
- What’s really holding enterprise AI back?, MIT Technology Review Insights
- Apache Ossie (incubating), Apache Software Foundation
- AI Registry, GitHub
- Analytics playbook, GitHub
- What is Atlan Trident, Atlan Trident documentation
- Bob Muglia profile, Forbes
- Josh Klahr author page, Snowflake
- Context Conference 2026 sessions and transcripts, Atlan
- Recap of the day, Context Conference 2026
- Inside Verizon’s context architecture, Context Conference 2026
- The company brain for shared skills and ontologies, Context Conference 2026
- The Post-AI Data Stack, Context Conference 2026
- Will open context interchange ever be real?, Context Conference 2026
