AI Agents do all my work
TL;DR
Andrew Wilkinson is already running slices of a real business with AI agents — his startup Deep Personality has support, dev, and marketing agents in Harbor/OpenClaw handling tickets, shipping PRs, and managing Meta and Reddit ads, with about $20,000 in revenue so far.
The magic isn’t “fully autonomous company” yet — it’s highly scripted genius interns — Wilkinson says today’s agents are like a “genius baby” or “Zapier zaps that can make basic decisions,” useful when carefully trained but nowhere near true CEO-level autonomy on April 29, 2026.
His biggest edge comes from feeding models context, not just prompts — he records meetings with Fireflies, ingests email and daily-life transcripts into GBrain, and builds vector databases for Tiny and his family office so he can query things like 132 venture investments, $16M invested, and $36M current value.
AI let his CFO build software that would have been absurd to custom-build a few years ago — instead of paying Addepar $50,000 to $100,000 a year, Wilkinson’s non-technical CFO vibe-coded a live portfolio dashboard and AI assistant in roughly two weeks.
He thinks software moats are collapsing fast — Wilkinson compares today’s software market to restaurants and “cigar butts,” arguing that AI makes products easier to copy, pushes pricing down, and shifts value toward distribution, credibility, or hard-asset bets like TSMC and data centers.
Some of his most compelling uses are intensely personal, not corporate — he built agents that draft email replies as multiple choice in Telegram, generate a custom 7-minute-style daily news podcast, and monitor Apple Health data closely enough to flag viral nerve-pain flare-ups from wrist-temperature changes.
The Breakdown
The OpenClaw high and the Arizona moment
Wilkinson opens with pure addict energy: after AI coding tools felt “interesting but not there yet,” he says by late 2025 he was waking up at 3 or 4 a.m. to sit in Claude Code with 10 tabs open. The breakthrough came when he forgot his laptop on a trip to Arizona and still ran his business from Ubers through an OpenClaw agent — a “chasing the dragon” moment because nobody noticed his emails were AI-written.
Why he built Deep Personality instead of hiring people
The product started from a very human experiment: he and his girlfriend took psychological tests, dropped the JSON into ChatGPT, and were stunned when it accurately mapped their fights and relationship dynamics. That became Deep Personality, a 40-minute test that generates a 100-page Robert Greene-style report covering archetypes, attachment style, ADHD/OCD signals, job fit, and relationship patterns.
The anti-org-chart startup: agents for support, dev, and marketing
Rather than staffing up, Wilkinson used Harbor — a friend’s GUI layer for agent workflows — to run the business with dev, support, and marketing agents. The support agent can solve tickets, route bugs to dev, and even auto-merge PRs for P0 issues; the marketing agent plugs into PostHog plus Meta and Reddit ads to test creative, set budgets, and execute campaigns, with Wilkinson wondering what happens when he hands it a $100,000/month ad budget.
What AI company autonomy can and can’t do right now
Both he and Greg push back on the hype around “autonomous companies.” Wilkinson’s blunt version: current agents are not CEOs, they’re “a genius baby” you still have to teach every step, while Greg says anyone selling fully beach-proof autonomy today is mostly selling a dream.
Vector databases as the eye of Sauron for messy businesses
A big chunk of Wilkinson’s setup is about memory: Fireflies records meetings, nightly cron jobs turn them into markdown, and GBrain/vector databases make everything searchable for models. In his family office, that lets him ask one question and instantly get a breakdown like 132 direct investments, $16 million invested, $36 million current value; in Tiny’s 24-business sprawl, it helps surface “icebergs,” headcount questions, and weird trends that would be impossible for one human to hold in their head.
His bearish take on software moats
This is where the mood turns from thrilling to slightly grim. Wilkinson argues software has gone from scarce and defensible to cheap, copyable, and crowded — his funeral-home analogy is that every operator now has an AI-powered “web whiz nephew,” so niche software gets built everywhere and prices collapse; his answer for ambitious young builders is basically: make $1–2 million fast, then maybe buy TSMC or data center exposure.
The family office runs on Claude spend instead of payroll
At his holding company, he says they’ve scaled API costs instead of headcount, with a roughly $40,000/month Claude bill doing work that might otherwise require more employees. The cleanest example is finance: his CFO, who had never coded, built a custom portfolio and balance-sheet system in two weeks to replace expensive tools like Addepar, complete with live holdings, risk analysis, and a data-aware AI assistant.
Personal agents for life admin, media, and health
The final stretch is the most sci-fi: Wilkinson records his whole day with a vibe-coded app called Hearsay, ingests it into GBrain, and uses agents to triage email, infer projects, and send him multiple-choice response options in Telegram. He also built a custom daily audio briefing from newsletters and Readwise using Gemini voice, plus a health agent that watches Apple Health metrics, medication adherence, and even spotted a pattern where wrist temperature shifts three days before his nerve pain flares.
His favorite prompt trick: let the model interview you
Wilkinson closes with one practical tip he says changed everything: don’t write the perfect prompt yourself — tell Claude your goal and ask it to “ask me a shitload of questions” using multiple-choice tools until it builds the prompt for you. He also recommends forcing a team of eight sub-agents on harder tasks, because the quality jump has been dramatic in his own workflow.
Share