Domain-Driven Agents

The author argues that LLMs struggle with legacy codebases because they lack understanding of domain-specific context and inconsistent naming conventions, rather than due to technical limitations. To address this, the author proposes splitting work into strategic (human decision-making about design) and tactical (LLM-assisted mechanical implementation) components, allowing humans to focus on the high-value work of clarifying system architecture while LLMs handle code refactoring at dramatically reduced cost. This approach makes paying down technical debt more feasible by automating the typing while preserving human judgment on what should change.