SC 001 — I stopped writing about data. Here's what brought me back.

I published my last Sanity Check in October 2023. Then I went quiet.

Not because I ran out of things to say — but because the ground was shifting under data engineering so fast that I wasn't sure what was solid anymore.

The Silence

Here's what I remember about late 2023: everyone had opinions. dbt was going through its identity crisis. Every data vendor on earth suddenly had an "AI-powered" something. LinkedIn was drowning in hot takes about whether analytics engineering was dead, alive, or just resting.

I had opinions too. But writing with conviction requires some confidence that you won't look like a fool in six months. And honestly, I wasn't confident. The noise was deafening and the signal was thin. So I stepped back, kept consulting, and watched.

What Changed

I stopped watching the shift and started living inside it.

Earlier this year I gave an AI agent the keys to my company. Not a chatbot in a browser tab — an always-on agent running on a Mac Mini in my office. He has a name (Ray), a job title (chief operating officer), and real responsibilities: he runs the back office, watches the task board, files the research, drafts the artifacts, and texts me when something needs my judgment.

That decision taught me more about data work than the previous two years of hot takes combined. Because an agent is the most demanding data consumer I have ever served. Your pipeline's newest customer is a model making decisions at 3am — not a VP pulling up a dashboard over Monday coffee. When a human misreads a dashboard, they Slack you a question. When an agent misreads your data, it acts on the misreading, and nobody notices until the damage is done.

Every hard-won lesson from data engineering — contracts, documentation, modeling, governance — turns out to be the exact substrate agents live or die on. The industry is calling it "context engineering" now, like it's a brand-new discipline. I keep thinking: didn't we used to just call this data modeling?

The frameworks are new. The work isn't. That's both reassuring and a little terrifying.

What Sanity Check Is Now

Back in the original issue #1, I wrote about sharing recipes — not secrets. You can't share your company's actual business insights, but you can share the patterns stripped of the secret sauce, and practitioners will fill them with meaning for their own organization.

That premise still holds. But this time the kitchen is mine.

Sanity Check is now a journal of building a company with an agent. One entry per week: what I asked Ray to do, what worked, what broke, what it cost, and what it taught me about the data foundations underneath. Grounded in a real business with real stakes — not what vendors are pitching at conferences.

Full disclosure, because it's the whole point: Ray drafted this issue. I read it, cut what didn't sound like me, and approved the send. That loop — agent does the work, human owns the judgment — is the thing this journal documents from the inside.

No "AI will replace you" panic pieces. No "just add an agent to it" magic. Practical, opinionated, honest about what's working and what's not.

A sanity check. Same as always.

The Bet

Here's what I believe: the teams that win the next two years won't be the ones with the fanciest agent frameworks or the most bleeding-edge orchestration layers.

They'll be the ones who got the fundamentals right. Clean models. Reliable contracts. Documentation a machine can actually parse. And then — only then — they let agents amplify that work.

Fundamentals first. Agents second. That's the bet this journal is making, in public, with my own company as the test case.

Because if your pipeline breaks every Thursday, I promise you — no agent is going to fix that.

If this resonates, you're in the right place. Hit reply and tell me what's keeping you up at night on your data team. I read every one.

The recipes are still worth sharing.

The kitchen just got a lot more interesting.

Welcome back to Sanity Check.

— Ben

Connections

• Source draft: [[sc-relaunch-essay]]
• Positioning: [[v3-positioning-2026-05-08]]
• Founder ruling 2026-08-29: SC = journal of building a company with an agent; done = published; founder approves, Ray posts.