Video cover

In this interview, Matt Pocock and Nate Herk discuss how to communicate intent to AI agents, improve the codebases they work in, and develop trust through observable results. Pocock describes a failed spec-driven workflow, the value of shared domain language, and opportunities in adapting agents to organizational knowledge work. The discussion offers personal experience and recommendations rather than controlled benchmarks. 14:07 24:03 28:02

Topic breakdown

  1. From skepticism to agent workflows
  2. Tell agents what you care about
  3. Learn with agents and primary sources
  4. Opportunities in knowledge work
  5. Expose unknown unknowns with Grill Me
  6. Stand out through communication and shared language
  7. Why Pocock abandoned his spec-driven workflow
  8. Models, effort levels, and application feedback
  9. Get started and climb the trust ladder
  10. What's next for Matt

Key takeaways

  • Explicit priorities remain necessary as models improve. Pocock argues that agents cannot infer what you care about, so skills and other processes must communicate it. 29:44 30:07
  • A maintainable codebase is part of the agent's working environment. After months of deteriorating results with a large-spec workflow, Pocock returned his focus to code that was easy to change and reported renewed progress. 28:02 28:22 28:46 29:08
  • Questioning reveals missing requirements. Grill Me interviews the builder to expose priorities and concerns such as security and deployment before the agent fills the gaps itself. 17:01 17:34
  • Shared product vocabulary reduces ambiguity. Pocock recommends using the same meaning for an order in client conversations, code, the database, and the UI. 24:03 24:23
  • Better application access improved his agent's work. Enabling browser interaction, screenshots, and a functioning local application gave the agent better evidence and, in his account, substantially improved quality. 33:29
  • Reusable knowledge-work processes are a career opportunity in Pocock's view because verification is difficult and each organization works differently. 13:30 14:07
  • Start with boring, inspectable tasks and earn trust gradually. Pocock favors work with measurable outputs and easy corrections, while keeping humans involved as agent use expands. 37:54 38:09 40:04 41:03