How a bank's top team got tired of AI coding agents

A team from a Norwegian bank that practices mob programming and TDD describes their 5-month experiment with Claude for all coding. They concluded that for their brownfield, domain-heavy systems, AI agents were simply not fast or accurate enough. The payoff from context-switching, waiting for responses, and manual

review was lower than just writing the code themselves. The hype around AI 'solving coding' ignores the complexity of legacy critical infrastructure. The team found LLMs excellent for analysis, dashboards, telemetry, and prototyping. The pain point is the 'sugar rush' effect: initial excitement leads to fatigue from

constant waiting and review. They explicitly state that AI-generated code reduces their understanding of the domain and breaks their flow. Their core finding is that the best use of LLMs is as an accelerator on well-architected codebases, not a replacement for human coding. The paper they mention

(published in the Journal of Systems and Software) details how to introduce new ways of working via low-commitment experiments. This is a practical, evidence-based counterpoint to the 'agents for everything' narrative. The tradeoff is clear: speed on simple tasks vs. quality and maintainability on complex systems.