Make priorities explicit

  • Pocock argues that the human-agent communication gap persists regardless of model intelligence. Explain the desired outcome and the qualities you care about rather than assuming the agent will infer them. 4:24
  • The codebase influences subsequent output. In his account, a neglected codebase encourages more poor code, while caring about the working environment improves what agents can produce. 5:02

Raise the minimum quality of work

  1. Think like someone managing 100 developers: ask what would improve the work of the weakest contributor. This is Pocock's analogy for designing an agent's environment. 5:27
  2. Learn established engineering practices, especially how to write useful tests and checks that prevent recurring classes of bugs. Pocock sees these older ideas as directly applicable to agents. 5:56 6:16
  3. Make those quality priorities part of the conversation. Agents can help teach the practices when you ask informed questions and show interest in learning them. 6:37