Moltbot Creator Interview: How to Succeed at Agentic Coding

Why 'Closing the Loop' Separates AI Winners From Frustrated Coders

Peter Steinberger built PSPDFKit—the PDF framework on over a billion devices—then burned out and disappeared from tech for three years. When he returned in 2024, he skipped past GitHub Copilot and GPT-3.5 entirely, landing directly on Claude Code. Within months, he was merging 600 commits in a single day. He's also the creator of Moltbot (formerly Claudebot)—the open-source personal AI assistant that's taken the solopreneur world by storm.

On the slot machine psychology: "It's the same economics as a casino. You press the trigger and ding ding ding—it does crap or it blows your mind." The variable reward loop of AI coding is genuinely addictive when you learn to work with it rather than against it.

On what separates effective AI coding: "The good thing about how to be effective with coding agents is always: you have to close the loop. It needs to be able to debug and test itself. That's the big secret." This is why AI excels at coding but struggles with creative writing—code can be compiled, tested, and verified.

On not reading most of his code: Peter confesses he ships code he hasn't fully read, but explains why that's not as reckless as it sounds. "Most apps are just data coming in one form, packaged into a different form, stored in a database in a different form... We are pretty chasing printers. The hard part was solved by Postgres 30 years ago." He focuses on system architecture while letting agents handle the plumbing.

On PRs becoming "Prompt Requests": "I read the prompts more than I read the code because this gives me more idea about the output. I ask people to please add the prompts." The thinking about how it should work matters more than the generated implementation.

On running agents in parallel: He typically runs 5-10 Codex agents simultaneously, treating coding like managing a team of imperfect but brilliant engineers. When one agent takes 40 minutes to build a feature, he moves to something else: "I constantly jump around. Usually there's one main project that has my focus and satellite projects that need attention."

On why companies struggle with AI adoption: "I could easily run a company with 30% of the people. But this requires completely redefining how the company works... You have to do a big refactor first—not just on your codebase, but also on your company."

5 Key Insights for AI-Powered Developers

• Close the loop — Design systems so agents can verify their own work through tests, linting, and CLI tools. This is the fundamental principle.
• Think in architecture, not lines — You care about system structure; agents handle implementation details. "I'm the architect, Codex does the line-by-line understanding."
• Build CLI-first for everything — CLIs let agents filter, chain, and script operations—far more powerful than MCPs which can't be composed.
• Let go of perfectionism — "I used to obsess over every spacing, every naming. In retrospect—what the heck? The customer doesn't see the insides."
• Treat prompts as documentation — When reviewing PRs, the prompt someone used reveals more about solution quality than the generated code.

What This Means for AI-First Development Teams

Peter's approach represents a fundamental shift in how senior engineers might work in the AI era. Instead of writing code, they become architects who design systems for agent verification and taste-check the outputs. The skill isn't coding—it's knowing what questions to ask and how to structure feedback loops that make AI succeed.

For organizations considering AI transformation, his warning is clear: you can't just add AI tools to existing workflows. You need to restructure how the company thinks about code ownership, reviews, and what makes an engineer valuable.

Related: See how Moltbot works in practice: Moltbot Explained: Your 24/7 AI Employee