The Pragmatic Engineer — selected conversations
The creator of OpenClaw: "I ship code I don't read"
Eighteen curated conversations on agent workflows, design, harnesses, and software quality
A manually reviewed selection of Pragmatic Engineer conversations with Maggie Appleton, Matt Pocock, Tibo Sottiaux, Martin Fowler, Dex Horthy, Boris Cherny, Mario Zechner and Armin Ronacher, Laura Tacho, Hillel Wayne, Mitchell Hashimoto, Kent Beck, Peter Steinberger, Grady Booch, Steve Yegge, DHH, Dax Raad, Charity Majors, and Addy Osmani. Topics include design prototyping, reusable skills, agent workflows, context and harness design, TDD and verification, architecture, measurement, accountability, and open-source governance. Practitioner observations, product-team claims, and forecasts are not controlled findings.
In this collection Browse 18 summaries 12 of 18
Gergely Orosz, host of The Pragmatic Engineer podcast, interviews Peter Steinberger, introduced as the creator of OpenClaw and founder of PSPDFKit. Steinberger describes reading less routine generated code while retaining responsibility for architecture, product feel, and executable verification; his throughput, staffing, and reliability claims remain personal reports from an unusually permissive open-source project 40:26–43:21(opens in a new tab), 1:07:17–1:08:20(opens in a new tab), 1:53:00(opens in a new tab).
Key Points Covered
- Unread code can still receive architectural care: Steinberger skips much routine plumbing but invests heavily in system structure, including a 15,000-line plug-in refactor 40:26–43:21(opens in a new tab).
- Shape the product through implementation: He uses partial outputs to discover alternatives and iterates on how a feature feels instead of relying on a fixed upfront specification 1:09:22–1:13:29(opens in a new tab).
- Parallel agents consume human attention: He reports using five to ten agents, but calls the workflow mentally taxing and organizes it around one main project plus smaller tasks 54:34–57:31(opens in a new tab).
- Close every executable loop: Agents compile, lint, test, and inspect output through CLIs and Docker-based end-to-end harnesses when normal interfaces are too slow or opaque 57:31–1:00:33(opens in a new tab).
- Testability shapes architecture: CLI-accessible core paths give agents faster feedback and push explicit reasoning about interfaces and validation 1:00:33–1:03:27(opens in a new tab).
- Intent can matter more than contributed code: He asks open-source contributors for prompts and detailed intent because reviewing small generated patches may cost more than rebuilding them in maintainer context 1:47:58–1:49:00(opens in a new tab).
- Speed and staffing claims are anecdotal: His roughly 600-commit day and 30%-of-staff estimate are personal claims, not controlled measurements; he says the required senior engineers would be hard to find 1:07:17–1:08:20(opens in a new tab), 1:42:51–1:43:55(opens in a new tab).
- Security limits generalization: Steinberger describes broad machine access, public Discord exposure, and a proactive heartbeat as risky; Orosz calls this more of a “YOLO project” than most production software 1:20:08–1:20:56(opens in a new tab), 1:25:38–1:26:41(opens in a new tab), 1:53:00(opens in a new tab).
Full video: https://www.youtube.com/watch?v=8lF7HmQ_RgY(opens in a new tab)