David Ondrej Podcast — selected agentic engineering conversations
Harness Engineering is the future… trust me.
Six curated interviews on software factories, agent environments, review, and harness design
A manually reviewed selection from the David Ondrej Podcast featuring Armin Ronacher, Dexter Horthy, Thorsten Ball, Matt Pocock, Kun, and Magnus Müller. The conversations examine how agentic engineering shifts attention from code generation toward product judgment, architecture, working environments, context, orchestration, verification, and small adaptable harnesses. Practitioner observations and forecasts are not controlled findings.
In this collection Browse 6 summaries 6 of 6
David Ondrej interviews Magnus Müller, co-founder and CEO of Browser Use, about thin, self-modifiable browser harnesses and an emerging interaction model in which always-running agents propose actions against high-level goals while people retain approval and responsibility.
Key Points Covered
- The interface becomes the problem: Müller says reliable browser control does not by itself determine what people should ask increasingly capable agents to do. His experiment reverses the normal prompt flow by having the agent suggest work against a high-level goal 02:36–04:19(opens in a new tab).
- Proactive agents turn prompting into approval: His Telegram setup connects work services, checks for useful actions, and presents concise proposals that a person can approve, reject, or refine 05:21–08:06(opens in a new tab).
- Agent-originated work changes ownership: After an agent independently pursued a naming issue, Müller describes feeling less like the initiator and more like an employee; he then asks the system to explain each proposal's likely impact so he has a reason to care 08:58–11:47(opens in a new tab).
- A small harness lets the agent create missing tools: Müller presents Browser Harness as roughly 600 lines that an agent can inspect and extend for uploads, cross-origin frames, drawing interactions, dialogs, and other browser edge cases 29:17–32:52(opens in a new tab).
- Better coding models enable thinner abstractions: He argues that older models needed larger abstractions and carefully written tool prompts, whereas newer coding agents can combine low-level functions and modify the harness directly 32:52–33:55(opens in a new tab).
- Self-modification creates an update problem: When each running agent can alter its own source, local versions diverge from upstream. Müller proposes letting the agent choose how to reconcile changes and fix defects in the same harness it operates 25:29–27:22(opens in a new tab).
- Human approval is both control and bottleneck: The discussion distinguishes private autonomous work from externally visible actions such as emails or pull requests, where Müller still wants approval even though repeated swiping limits throughput 41:12–43:14(opens in a new tab).
- Engagement and agency remain unresolved: Müller worries that excessive suggestions become ignorable and that optimizing proposals for acceptance or addictiveness may increase productivity while reducing the person's sense of ownership 36:21–41:12(opens in a new tab).
Full video: https://www.youtube.com/watch?v=k8o_UgtrGyA(opens in a new tab)