Anthropic Proved the Agentic Data Stack Works.

I just read Anthropic's post on self-service analytics. Nothing surprised me. It described, almost line for line, the bet I made when I started building OptimaFlo.

Their core claim is blunt: writing SQL is trivial now. For people, and for the model. The hard part is everything wrapped around it.

Which table should the agent use? Which definition of "revenue" is the real one? Which dimensions are safe to join? Which source went stale last week? When someone asks for "active users," what counts as active, and does it match what the last person who asked meant?

Point an agent straight at your warehouse and you find out the hard way. It feels like magic for a little bit but then it starts handing over confident, wrong answers, and nobody catches them until the number is already in use.

Anthropic's fix isn't as loud and groundbreaking as people thought. Canonical datasets. A semantic layer. Curated reference docs. Lineage. Validation. Skills that encode how a good analyst actually works. Their agent went from under 21% accuracy to consistently over 95%, and the jump came from building context and tools around the model.

The detail I keep thinking about

They gave the agent direct access to thousands of their own past queries. Every question, already answered correctly, sitting right there. Accuracy barely moved, about less than one point. The agent read the queries and still did not use them.

The bottleneck was never access to data. It was structure: taking a messy human question and mapping it to the one correct entity, then knowing how to work with it. That is the part no model solves for you just by getting bigger. I have watched it play out in our own product. You can hand the model everything and still get a shrug, because "everything" is not the same as "the right thing, defined."

I agree with all of it. I have believed it for over a year.

Who builds the scaffolding

Every layer Anthropic describes is built and owned by a data team. The governed models. The semantic layer. The colocated repo. The eval harness. The maintenance hooks that keep docs from rotting. All of it is the day job of people who do data for a living. But Anthropic has one of the best teams in the world.

A lot of companies that want self-service analytics do not have that. They have a small team or even one data person (oftentimes frustrated and overwhelmed). Sometimes zero, and a founder running SQL at midnight.

So the post raises the bar in a way that is easy to miss. "Point an agent at the warehouse" does not work. "Build and maintain the full agentic stack" works, but it needs the exact team you were trying to get by without. That is the gap OptimaFlo exists to close.

Same scaffolding, one person or a small team to run it

My bet is simple. The structure Anthropic describes should not be a project you staff. It should come with the product.

The canonical models, the semantic layer, the reference docs, the validation step: we build that for you, in your own cloud, on open formats like Iceberg, so it is your data the whole way through. One data owner or a small team can run the stack a whole team used to run.

Anthropic said it as plainly as I would; "We recommend generating the documentation with Claude, but having a human own the definition." That's it. Someone still has to decide what "revenue" means. What "active" means. What counts as retention. The model can't make that call and it shouldn't try.

That is why every answer in OptimaFlo passes a human approval step before it ships. Not as a compliance checkbox. The data owner is the one who knows the business. The agent should do the grunt work of finding, joining, and checking. We amplify the owner. We do not replace them.

So here is the real question

If you are a data owner reading Anthropic's post and feeling the weight of that to-do list, canonical datasets, semantic layer, evals, maintenance, all of it, that is the right instinct. It is a lot. It is also exactly the work we think you should not do by hand.

Anthropic just made the case that the agentic data stack is table stakes. They are right.

The question they leave open is the one I care about: who builds and maintains it when you do not have a full data team?

If that is you, I would like to know your plan. Come tell me, or see how OptimaFlo runs the stack for one owner.

---

Anthropic's post: "How Anthropic enables self-service data analytics with Claude" (June 3, 2026).