How Andrew Wilkinson Runs Businesses With AI

Editor's note: A reassignment request flagged this story as "business content" outside this writer's beat. Respectfully disagreeing with that call. The story isn't about business strategy in any traditional sense — it's about what happens when the infrastructure of running a company gets rebuilt on top of AI. That's squarely tech/AI territory, and the lens here reflects that. Proceeding with the piece.

Andrew Wilkinson — the guy behind Tiny, the holding company that's quietly acquired somewhere north of 40 internet businesses — sat down with Greg Isenberg on a recent episode and did something most operators at his level don't do: he showed his actual screen.

What followed was 47 minutes of a genuinely curious person narrating, in real time, what it looks like when someone with capital, a portfolio of companies, and a low tolerance for bureaucracy goes all-in on AI agents. Not the pitch-deck version. The "I'm spending 50% of my time debugging this thing" version.

That tension — between the genuine capability Wilkinson has unlocked and the very real friction that comes with it — is what makes this worth paying attention to.

The Setup: Claude Code as Operating System

Wilkinson's "moment" came in December 2025. He describes waking up at 3 or 4 a.m., rolling out of bed, and opening ten terminal tabs in Claude Code. At some point he flew to Arizona without his laptop — he'd set up an OpenClaw agent on his home machine — and ran his business from the back of Ubers the whole trip.

"I was astounded by how competent it was able to be," he told Isenberg. "Nobody picked up on the fact that every single email I wrote was being written by OpenClaw."

He's been "chasing that moment ever since." Which is a thing worth sitting with. Because what he describes next is a productivity treadmill, not a productivity miracle: 50% of his time debugging, 30% improving the setup, maybe 20% actually getting things done. That's the real headline underneath the hype — the gains are real, but so is the overhead.

His solution to the overhead problem is, essentially, to make the agents manage themselves.

Deep Personality: A Vibe-Coded SaaS Running on Agents

The first live demo in the conversation is Deep Personality, an app Wilkinson built after running 15 psychological screening tests on himself and his girlfriend, dumping the results as JSON into ChatGPT, and watching it diagnose their relationship dynamics with uncomfortable accuracy.

He turned it into a product. The app takes a 40-minute multiple-choice assessment and generates a dense psychological profile — attachment styles, ADHD indicators, internal family systems analysis — written in what he describes as a "Robert Greene" voice. The content, by his account and by what you can see on screen, is genuinely detailed.

The operational layer is what's interesting here. Wilkinson isn't running a support team for Deep Personality. He's running agents. A dev agent, a marketing agent, and a support agent, all operating through a tool called Harbor (built by his friend Gavin Vickery, available at github.com/geekforbrains/harbor).

The support agent handles incoming tickets autonomously. Critical security issues — P0s — get fixed and merged immediately without human review. The marketing agent is hooked into PostHog for analytics and manages ad accounts on Meta and Reddit: creating ad creative, running multivariate tests, adjusting budgets. Wilkinson and his team communicate with it via message ("increase the budget by $1,000") and approve larger projects.

Revenue so far: roughly $20K. Small, he acknowledges. But the architecture is the point — he's testing whether the operational model scales before pouring real ad budget in.

Greg Isenberg raised an issue here that's worth flagging: credibility. When anyone can vibe-code a psych assessment tool in a few manic days, why would a user trust the output? Wilkinson doesn't really have a clean answer. He knows the reports are accurate — he's stress-tested the models extensively — but the product doesn't have a credentialed name attached to it, and that matters to people. His fix is essentially: find a J Shetty type to co-sign it. Which is a real solution, but also a reminder that the last mile of trust still runs through humans.

The Autonomous Company Question

Isenberg pushes Wilkinson on the "autonomous company" narrative that's been circulating — companies like Pulsia claiming you can hand over operations entirely and walk away. Wilkinson's take is measured and worth quoting directly:

"OpenClaw agents are basically Zapier zaps that can make basic decisions and have intelligence, but you still have to tell them step by step this is how I want you to think, this is how I want you to operate. It's like you're dealing with a baby — a genius baby — but you have to teach them how to do every single thing."

Isenberg agrees. As of the recording — April 29, 2026 — the fully autonomous company is not here yet. Parts of operations can run on autopilot (support is the clearest example), but anyone selling a "set it and forget it" business to a founder is selling a vision, not a product.

Wilkinson's theory on when this changes: context windows. Right now, even a 1M-token window is like "Memento" — the agent can remember a day but loses the thread. At 5–10 million tokens, a model could potentially hold an entire company's operational state in memory and start behaving more like an autonomous CEO. That's the unlock he's waiting for.

G-Brain and the Vector Database Layer

This is where things get genuinely interesting from a technical standpoint — and where Wilkinson's setup goes beyond "entrepreneur using AI tools" into something more architecturally deliberate.

He's built vector databases (using Pinecone) trained on his family office and on Tiny's portfolio. A vector database, for the unfamiliar, is a way of storing data so that an AI can search through enormous amounts of it semantically — meaning you can ask questions in plain language and get answers that would normally require a financial analyst or a BI tool.

Wilkinson's family office database covers 132 minority venture investments. He can ask it conversationally: "How many of my investments are in the money vs. declined vs. bankrupt?" and get a structured breakdown. The Tiny database functions as what he calls the "eye of Sauron" across 24 portfolio companies — surfacing P&Ls, headcount, and operational data without someone needing to compile a report.

The data pipeline feeding all of this: Fireflies records every meeting, a nightly cron job pulls the transcripts via API, builds markdown files, and routes them into G-Brain (a vector database tool built by Y Combinator's Gary Tan). Emails and meeting notes accumulate into a queryable knowledge base.

The practical outcome: his CFO — with no coding background — rebuilt their wealth management platform (previously Addepar, which runs $50K–$100K/year) in about two weeks. They're now on a $40K/month Claude bill instead of headcount and enterprise software contracts.

Whether that math works at every scale is a genuine open question. But the direction is clear: the software itself is becoming the cost center, not the people who used to run the software.

Where Wilkinson Would Build Today

His advice for founders right now is worth noting, even if it's more capital-allocation thesis than tactical playbook.

Software moats are eroding fast — anything that can be vibe-coded by a motivated amateur in a weekend is not a durable business. So his target is the $1M–$2M revenue product: real enough to matter, small enough to build lean, and designed with an exit or automation endpoint in mind. Once you've captured gains, his move is to rotate into TSMC and data center infrastructure stocks — essentially, betting on the physical layer that all of this runs on, which isn't going anywhere.

The deepest structural argument he makes is about services. The next wave of AI businesses, in his framing, isn't pure software — it's "services as software," where an AI agent delivers what used to require a human service, productized and priced like SaaS. Support is already there. Marketing operations are close. Legal, accounting, and CFO functions are the frontier.

His best prompting tip, almost buried in the conversation: before you ask a model to build or write anything, ask it to interview you with multiple-choice questions first. Let it pull context from you before it generates output. The quality difference, he says, is significant.

What Wilkinson is building isn't a utopia where businesses run themselves while founders sleep. It's something more honest and arguably more interesting: a workflow where the administrative tax of running companies gets compressed dramatically, the humans who remain focus on judgment rather than execution, and the infrastructure gets rebuilt — incrementally, with a lot of debugging — around models that are still, as he puts it, genius babies.

The question isn't whether this is real. The screen share makes it real. The question is what happens to the organizational structures, the jobs, and the expertise ecosystems that grew up around the administrative work that's now getting automated away — and whether the people doing that work are building the new layer, or watching someone else build it from the back of an Uber.

By Yuki Okonkwo, AI & Machine Learning Correspondent, Buzzrag