
Promising LLM workflows and legacy code

How AI will change software engineering – with Martin Fowler
Topic 7 of 23
Published Nov 19, 2025 · Clip 26:27–33:39 (7:12)
Martin highlights two clear success areas for LLMs: rapid prototyping/explorations with throwaway code, and understanding existing legacy systems, where Thoughtworks colleagues populate a graph database from code semantics and interrogate it RAG-style — a technique strong enough to be in the Adopt ring of the radar. He notes areas still being figured out: working with an LLM one-on-one to build quality software using thin rapid slices, treating each slice as a PR from a productive but untrustworthy collaborator, whether LLMs can safely modify legacy code (illustrated by James Lewis burning tokens renaming a class that IDE refactoring already solved), and how teams of people should work with AI since demand for software means teams will persist.