Medium: Building a company with AI agents after 15 years in product management

What the article is about

Robert Cowherd spent a decade as a software engineer and fifteen years in product management at Uber, Amazon, and several startups before testing whether AI agents could replace not just individual tasks, but an entire cross-functional team. The experiment produced a complete product — 846 git commits — built solo over four weeks, with agents handling engineering, marketing, sales, customer support, and design.

Context

The article is structured as a set of operating principles for anyone attempting a similar approach. Cowherd was not automating a single repeatable task but designing an agentic operating system with orchestrator agents managing sub-agent teams in each functional area. That scope separates this from most writing on AI workflows, which tends to focus on incremental task substitution rather than wholesale team replacement.

Key method and takeaways

The most consistently applicable insight concerns the nature of direction. Cowherd found that stating the problem — not the solution — produced better agent output. Traditional PM work involves translating business goals into precise specifications; with agents, that precision becomes a constraint. Narrowing the output too early prevents agents from exploring alternatives that a human team might have surfaced through iteration.

A second principle addresses the psychological friction of delegation. The instinct to verify and correct every agent output — working alongside the system rather than through it — is described as the primary obstacle to building something that scales. Cowherd calls this “resisting manual execution” and frames it as a trust-building process: the system needs feedback loops, not intervention at each step.

The article also covers infrastructure decisions that most product teams will encounter as agent usage scales. Markdown files and shared knowledge documents become critical continuity infrastructure between sessions. Customer-facing agents — support, social media — create distinct security surfaces that require API-mediated access and least-privilege design. Traditional PRDs lose value when iteration cycles compress from weeks to hours; Cowherd argues it is better to build and test assumptions rapidly than to specify everything upfront.

Who it is useful for

Product managers building with AI agents — whether at a startup, inside a larger team experimenting with autonomous workflows, or as a solo PM trying to extend capacity without adding headcount. The practical framing and the author’s engineering background make the operational advice credible rather than speculative.