David Ondrej Podcast — selected agentic engineering conversations
Matt Pocock’s Agentic Engineering Workflow (just copy him)
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 4 of 6
David Ondrej interviews Matt Pocock of AI Hero about an agentic engineering workflow that delegates scoped implementation while keeping architecture, prioritization, product direction, and review under human control.
Key Points Covered
- Keep strategy human-owned: Pocock distinguishes tactical code production from strategic decisions about architecture, interfaces, tests, and task boundaries. AI can perform much of the former. Pocock says effective delegation still depends on the latter 00:00–02:52(opens in a new tab).
- Use skills as deliberate procedures: Pocock prefers user-invoked procedures such as Grill Me over automatically invoked abilities. This keeps the developer in control and avoids filling context with every skill description 17:00–21:51(opens in a new tab).
- Turn experience into reusable team workflows: Pocock says senior developers can package repeated planning and engineering processes as skills. Teammates can use and improve them. However, Pocock distinguishes reusable knowledge and skill from context-dependent wisdom 21:51–24:02(opens in a new tab).
- Isolate work that runs while you are away: His AFK (away-from-keyboard) setup combines local Claude Code planning with Sandcastle sandboxes and GitHub Actions. Agents can implement and review separate tasks asynchronously while the developer is elsewhere 24:51–26:59(opens in a new tab).
- Optimize the harness as well as the model: Pocock argues that prompts, skills, codebase architecture, tests, and feedback systems deserve as much attention as model choice. Developers can directly improve those conditions 27:47–32:34(opens in a new tab).
- Delegate tactics without delegating direction: Better models can complete broader scoped tasks, but Pocock says humans still need to set goals, judge whether results are correct, and retain product and roadmap decisions 33:38–36:49(opens in a new tab).
- Fix the system behind each defect: When a model finds a bug, the durable response is to strengthen recurring tests, security checks, or review processes rather than only patching that instance 38:54–41:50(opens in a new tab).
- Think in queues, not endless loops: He treats AFK development as multiple agents taking scoped tasks from a prioritized queue, with exploration, implementation, review, and merge represented as explicit stages 42:53–48:08(opens in a new tab).
- Preserve observability while reducing review cost: Pocock says human checkpoints can move later for lower-risk work. But maintainers still need sampled review and rich evidence. They need both to assess generated code and the system producing it 48:08–53:09(opens in a new tab).
- Rebuild the setup from a baseline: Pocock recommends removing accumulated skills, plugins, MCP servers, and instruction files, observing the uncustomized agent, then adding back only procedures that address demonstrated needs 1:00:05–1:02:02(opens in a new tab).
Full video: https://www.youtube.com/watch?v=nQwJVHCtDDY(opens in a new tab)