Descriptions:
Peter Yang interviews Matt Van Horn, a self-described non-technical founder who has built a reputation as one of the most prolific AI product builders working today. The conversation centers on Matt’s “agentic engineering” system — a practical, evolving workflow for shipping real software products without writing or reading code — which went viral on X and forms the basis of this deep-dive.
The cornerstone of Matt’s approach is what he calls “compound engineering” using Claude Code’s `/plan` and `/work` commands. He argues that agents are inherently lazy and tend to optimize for fast user approval over thoroughness, so forcing them to write a structured step-by-step plan before executing dramatically improves output quality. Crucially, Matt has stopped reading those plans himself over the past two months, instead asking targeted natural-language questions to verify intent before letting agents run. He demonstrates this live with “Agent Cookie” — a Tailscale-based tool that syncs browser cookies across devices so AI agents (Vercel’s AI browser, browser-use, Playwright) stay authenticated — and “The Printing Press,” a system that generates command-line interfaces for any web service by sniffing HAR traffic for hidden APIs. The library has grown from roughly 40 CLIs at launch to over 200 community-printed ones.
The interview also covers the practical tradeoffs between CLIs and MCPs for agent authentication, how to structure multi-project agentic tasks, and the broader philosophy of moving fast by trusting agents with autonomous execution. Useful for both technical developers and non-engineers exploring AI-driven product development.
📺 Source: Peter Yang · Published June 14, 2026
🏷️ Format: Interview
