Engineers shift to research roles in AI workflows

Andy Treidman of Theory Ventures observes that the software engineering role is transforming from execution-oriented to a research or scientist-like role due to the non-deterministic nature of AI-generated code. Engineers must now focus on customer empathy, architectural design, and defining constraints rather than writing code. He notes that proper specification is critical; he uses a control loop requiring 20 to 30 questions to be answered before engineering tickets are created to prevent agents from making incorrect assumptions about user experience.

Regarding cost, Treidman states that while inference costs are rising, the return on investment is positive for startups because they can ship roadmaps twice as quickly without hiring additional engineers. He suggests that well-specified tasks allow companies to route work to cheaper models like Luna or Sonnet, reserving the most intelligent models for complex problems. He advises founders to start by encapsulating repeated tasks into skills and ensuring infrastructure allows agents to run frictionlessly in the cloud.

Treidman argues that engineers must uplevel their skills to focus on intent and customer outcomes. The velocity gains from AI justify the inference costs for early-stage companies, as the ability to iterate rapidly on non-deterministic systems provides a competitive edge.