Andrew Wilkinson Runs SaaS & Life on AI Agents

Vibe-Coding: Pure Vision Without People

Andrew Wilkinson describes vibe-coding as a direct path from idea to execution, bypassing the 'middlemen' that dilute creativity in traditional business. He built Deep Personality, a personality assessment app generating detailed 100-page reports on archetypes, superpowers, kryptonite, ADHD/OCD likelihood, attachment styles, and career fit, after feeding psychological test results (15 screens, 40 minutes each) into Claude as JSON. The AI output stunned him and his girlfriend by precisely diagnosing their relationship fights. Rather than hiring via his agency Metalab, Andrew solo-built the app, design, and copy in 'a few manic days,' arguing 'the worst part about business is people'—employees misalign on vision, designers lack frontend skills, etc. "Creativity is just compromised based on how many people are between your vision and execution," he says. The app now earns ~$20K revenue, with agents handling the rest.

He acknowledges credibility challenges for vibe-coded products: without psychologist endorsement (e.g., partnering with Jay Shetty), tech-bro audiences hesitate to share. Still, viral newsletter posts prove demand, and he envisions celebrity couples endorsing it for trust.

Autonomous Operations: Harbor Agents as Org Chart

Andrew runs Deep Personality via Harbor (github.com/geekforbrains/Harbor), a GUI harness built by friend Gavin Vickery atop Claude, improving on OpenClaw's text-based limitations. Agents—dev, marketing, support—share a knowledge base with PostHog data, env vars, and docs. Support auto-resolves tickets or escalates P0 bugs (merging PRs instantly); marketing runs multivariate Meta/Reddit ad tests, creates creatives, adjusts budgets (e.g., +$1K on command), and plans SEO. "Support to me is not a job very very soon," Andrew asserts, while a $100K/mo ad budget could transform scaling.

In his family office, agents replaced headcount with a $40K/mo Claude bill. The non-technical CFO vibe-coded a net worth platform replacing Adapar ($50K–$100K/year) in two weeks. Andrew predicts 3-6 months until basic businesses hand off fully, but today's agents are 'Zapier zaps with intelligence'—genius babies needing step-by-step instructions. Full autonomy awaits 5-10M token contexts holding entire companies. He pushes back on Pulsar-like 'autonomous company' hype: "People selling that today are selling a dream."

"I was traveling... forgot my laptop... ran my entire business using OpenClaw in the back of Ubers... nobody picked up that every email was AI-written," recounts Andrew of his heroin-like high from Claude Code's December 2024 unlock, now tempered by 50% debugging time.

Data Centralization: Vector DBs as Company Oracle

Andrew's 'G-Brain' (Gary Tan's vibe-coded vector DB) ingests Fireflies transcripts, emails, and notes nightly, trained on Tiny (24 portfolio companies) and his holdings (132 minority investments). It conversationally queries P&Ls, headcount, and returns insights like '$16M invested → $36M value.' Claude Code is his OS; this setup acts as an 'eye of Sauron' across operations.

Centralizing pipelines enables agentic transitions: standardize data first, then route to agents. For Tiny and holdings, it surfaces trends without manual digging.

Future Bets: Services Over Software, Hardware Moats

Software moats erode fast; Andrew advises builders target $1-2M ARR products, then park gains in TSMC/data center stocks. "Services as the new software," he says, as agents commoditize rote SaaS tasks (support, accounting). The new interface: conversational UIs over apps. Build now, as AI lowers barriers.

Where he'd build: vibe-coded credibility layers for personal tools; agent-harnessed services. Prompt tip: "Ask the model to interview you with multiple-choice questions before generating output"—mimics Deep Personality's accuracy.

"I feel like I started shooting heroin... waking up at 3 or 4 a.m.... 10 terminal tabs open." - Andrew on Claude Code addiction.

"OpenClaw agents are still like Zapier zaps... a genius baby you have to teach every step." - Andrew on agent limits.

"Once context windows hit 5-10M tokens, we can run entire companies." - Andrew predicting agent CEOs.

Key Takeaways

• Vibe-code solo to execute vision purely; bypass people for speed.
• Use Harbor-like harnesses for dev/marketing/support agents sharing PostHog data.
• Centralize data nightly into vector DBs (e.g., G-Brain) for conversational queries.
• Debug tax is 50%; chase highs but expect treadmill.
• Support jobs vanish soon; marketing agents reshape ad budgets.
• Prompt models to 'interview' via multiple-choice first for precision.
• Target $1-2M products; invest in TSMC/data centers as software commoditizes.
• Services > SaaS; build credibility layers for vibe-coded apps.
• Autonomy needs 5-10M contexts; today's agents need hand-holding.