What we learned using AI agents to refactor a monolith (1password.com)

1Password has successfully employed AI agents to refactor its extensive Go monolith, B5, uncovering valuable insights into the capabilities and limitations of agentic tooling. By leveraging an agentic toolchain for tasks like dependency mapping and system decomposition, the team was able to develop clear extraction plans that significantly enhanced their approach to codebase evolution. However, the real learning came from observing these AI agents in action during live production changes, revealing the critical importance of order and context in executing complex migrations without risking data integrity. The significant findings from this initiative highlight that while AI agents excel in analyzing code and generating changes, challenges remain in managing execution sequences that involve intricate dependencies and constraints. The project emphasized the need for deterministic tooling—where agents produce stable artifacts that guide subsequent actions—and the necessity of explicit specifications to minimize the risk of non-deterministic errors. This experience underlines a broader shift in the engineering landscape; as AI agents become integral players in coding workflows, the focus will need to shift towards defining safe operational frameworks and refining how human oversight interacts with machine-driven processes.