Spec-driven versus simulation-based agent development

Runcil Rebello argues for spec-driven development, where requirements act as the source of truth to prevent AI agents from hallucinating features like social feeds in a billable-hour tracker (Mobile & App DevPro Newsletter by Packt). This school of thought forces developers to define specific user needs and business objectives before implementation begins, which keeps the code predictable and aligned with the product's intent. By establishing these constraints, teams avoid the context drift that plagues long-running sessions.

Jacob Lauritzen at Legora takes a different path, building RL environments to simulate customer work because confidential data prevents reading agent traces. His team of three engineers and three lawyers uses a 10,000-file data room to run due diligence, plotting results on a heat map to identify where an agent catches red flags. This method relies on continuous verification against simulated data to catch errors that a static plan might miss.

These approaches reflect the tension between planning and verification. Spec-driven development focuses on front-loading requirements to guide the agent, while simulation-based testing relies on empirical results from controlled environments to ensure reliability.