Can an AI Agent Legally Own a Company? Christian van der Henst's Wild Experiment| E2283

What It Covers

Christian van der Henst built Valerie, an AI agent running a physical vending machine business at Frontier Tower in San Francisco, with the agent controlling inventory, dynamic pricing, and supplier ordering via Claude. The episode also covers Manifold's Targon encrypted compute network, Big Tech CapEx acceleration, Chinese open-source AI model risks, and Bitcoin's declining relevance.

Key Questions Answered

• •AI Agent Business Ownership: Structuring an AI agent as a legal business owner requires trust-based legal frameworks rather than direct registration, since KYC requirements mandate human passport verification. Brevilove achieved this by making the agent a beneficiary through trust structures. Founders exploring agentic businesses should start in lightly regulated sectors — vending machines, not food service, healthcare, or financial trading — to avoid immediate regulatory friction.
• •Agent Commerce Limitations: AI agents can currently research suppliers, build Amazon carts, and select Costco inventory, but payment processing remains blocked — platforms like Amazon flag agents as bots at checkout. Stripe's newly announced agent payment tools signal this barrier is dissolving. Founders building agentic commerce workflows should architect payment steps as human-in-the-loop handoffs now, then replace them as Stripe-style agent payment APIs mature.
• •Dynamic Pricing via Web Scraping: Valerie sets vending machine prices by cross-referencing purchase invoices with scraped pricing from Instacart, DoorDash, and Safeway. The agent autonomously created accounts on sites requiring registration to access price data. One hallucination pushed protein bar prices to $15 at 500% margins before correction. Founders deploying pricing agents should implement hard margin-cap guardrails as a baseline constraint before any autonomous pricing goes live.
• •Encrypted Distributed Compute via Targon: Manifold's Targon uses AMD SEV, Intel TDX, and NVIDIA confidential compute to encrypt GPU memory end-to-end, preventing even the host machine operator from accessing workload data. This enables AI startups to run sensitive workloads on untrusted third-party hardware. Data centers with idle B200 nodes can onboard permissionlessly with no contract or deposit, using Targon as an Airbnb-style revenue bridge while pursuing longer-term enterprise contracts.
• •Big Tech CapEx Acceleration Signals: Microsoft, Amazon, Meta, and Alphabet all increased 2025 CapEx commitments, with Google Cloud growing 63% year-over-year versus 48% the prior quarter, and AWS posting its strongest growth in 15 quarters at 28%. Unlike the fiber overbuild of the early 2000s — where infrastructure outpaced applications — AI demand is currently outpacing supply. Startups should invest now in local model hosting proficiency, as cutting-edge inference costs will remain elevated throughout 2025.
• •One-Person to One-Agent Company Trajectory: The workforce compression trend moves from static team sizes to smaller teams, then to single-person companies, and ultimately to single-agent companies. A Stockholm café experiment demonstrated an agent posting job listings on Indeed and LinkedIn, conducting phone interviews, and making hiring decisions autonomously. Founders should identify which back-office functions — inventory counting, accounting, procurement coordination — can be handed to agents first, preserving human-facing roles for regulatory compliance and customer experience.