Episode Summary
Executive Summary: Peter Steinberger describes a shift from traditional software engineering to AI-agent-driven building: he now thinks in architecture and prompts, not line-by-line code or PRs. Drawing on PSPDFKit and his new Cloudbot project, he argues effective AI coding requires tight validation loops, fast iteration, and product taste, while predicting major changes to team structure, onboarding, and developer workflows.
Main Topics: From PSPDFKit to AI-native building: Peter recounts his path from rural Austria, early tinkering, and building PSPDFKit into a major PDF framework, then burning out and stepping away before returning with a radically different workflow centered on AI agents. Cloudbot and the hyper-personal assistant vision: He explains Cloudbot as a deeply personal assistant that can act across apps, remember context, and proactively help—closer to a real Siri-like companion than a simple task bot. Closing the loop as the core AI coding principle: Peter argues that AI coding works when the agent can compile, test, lint, and debug its own output; verification, not blind trust, is what makes the workflow viable. Prompts over pull requests: He says he now values prompts more than code diffs, treats PRs like 'prompt requests,' and prefers architectural discussion plus agent-driven implementation over conventional code review. Why AI changes software architecture and team structure: Peter believes AI reduces the need for large teams, shifts value toward high-agency builders, and forces products and companies to be designed for agent efficiency from the start. Lessons from burnout, craft, and support: He reflects on the stress of running a company, the loneliness of leadership, and how hands-on support, blog posts, and product polish helped PSPDFKit succeed.
Key Arguments: AI makes coding fast, but the real skill is steering agents through good architecture and validation loops. You should optimize systems for agents, not just humans, because the model executes the code and needs low-friction paths. PRs are less useful than the prompts that produced them; prompt quality reveals the real reasoning behind the output. Code reviews are shifting from detailed line review to higher-level judgment of structure, intent, and prompt quality. Most software is plumbing and data transformation; the hard part is architecture, interfaces, and making things feel magical. Large companies will struggle to adopt AI efficiently unless they rethink both their codebases and their organization around agentic workflows. The best AI workflows are interactive and conversational: discuss options first, then build, then let the agent test and fix itself. Good software still requires taste; AI speeds up iteration, but product feel and polish remain differentiators. New engineers can use AI as an infinitely patient tutor, but they must be curious and learn through building and probing existing systems. Cloudbot’s value comes from combining proactive context, multiple channels, and memory into one assistant that feels like a real companion.
Data Points: PSPDFKit scale: used on more than 1 billion devices - Peter’s prior company/framework reached massive distribution across PDF-rendering devices. First indie app revenue: about $10K in the first month - His early iPhone dating-app client, built by parsing HTML, quickly became a meaningful side business. Initial app price: $5 - The app was launched in the App Store at a low price point before taking off. PSPDFKit team size at sale: 70 people - Peter says the company had grown to about 70 when he sold his shares. Current PSPDFKit size: almost 200 people - He notes the company continued growing after his departure. Time away from tech: more than 3 years - He stepped away from software and needed time to decompress after burnout and exit. Cloudbot community growth: 100 stars to 3,300 stars in a week - He describes the project’s rapid open-source/community explosion. Merge volume in one day: 600 commits - Peter cites an example of extreme throughput enabled by AI-assisted workflows. PR volume: 500 pull requests merged already - He says Cloudbot generated a huge influx of community contributions that he was effectively merging nonstop. Agent parallelism: 5 to 10 agents in parallel - He explains his current workflow on Cloudbot uses multiple agents simultaneously. Developer trust gap: 82% vs 96% - A cited Sonar survey: 82% say AI helps them code faster, while 96% do not highly trust AI code accuracy. Sonar scale: 750 billion lines of code daily - Presented as context for automated code analysis and verification. Notion experimentation scale: over 300 experiments per quarter - Used in sponsor messaging to illustrate mature internal tooling at scale. Feature flagging volume: over 600 features behind feature flags - Used to show how large teams manage risk while shipping quickly.
Pivotal Quotes: "I would say I write better code now that I don't write code myself anymore." — Peter Steinberger: Peter explains how AI-driven architecture and validation have improved his output even though he writes less code by hand. "The secret to making AI system development work well is to design your system to close the loop and have the AI run the test." — Host summary of Peter's argument: The episode’s recap emphasizes Peter’s central engineering principle for reliable AI coding. "I see them more as prompt requests now." — Peter Steinberger: He describes how he now thinks about PRs: the prompt and intent matter more than the diff itself.
Implications: AI is shifting software work from manual coding to system design, prompting, and verification. Teams that adapt workflows, architecture, and culture for agents may move far faster; those that don’t may be outpaced and may need fewer, more senior builders.