How 20 Practitioners Actually Build SE Agents: A Mixed-Methods Study of the Undocumented Craft

arXiv·medium signal

While benchmarks and deployment studies are common, little is documented about how developers actually construct software-engineering agents. Through semi-structured interviews with 20 practitioners across 12 organizations plus an online survey, this study surfaces the real design decisions, patterns, and pain points behind building LLM agents that navigate large codebases with limited human intervention. For builders, it's a rare grounded look at what working teams converge on versus what papers assume.