Fragments: July 21

Martin Fowler wraps up notes from the second Future of Software Development Retreat, highlighting key findings from a Thoughtworks report: code generation is no longer the bottleneck, verification is; 'harness engineering' is emerging; an apprenticeship crisis exists; executive/engineer expectation gaps pose risks; and legacy modernization offers near-term value. He explores the mismatch between engineers and C-suite views on LLMs, warning about security risks and board-level overconfidence. A cautionary tale illustrates how AI models applied in wrong contexts (e.g., desert-trained air filter optimization used in arctic) can lead to catastrophic failures. The article also addresses vibe-coding dangers, citizen developer risks, and the need for controls to manage shadow IT, emphasizing context, feedback sensors, and threat modeling.