Type: url
Original: https://arxiv.org/pdf/2604.26275
Status: generated
Summary
This single-author preprint from Northeastern University synthesizes industrial and academic work on agentic coding systems such as Claude Code, Codex CLI, Jules, Devin, OpenHands and SWE-agent, which it characterizes as operating at the granularity of a repository, feature or algorithm rather than a line or function. It proposes a six-layer reference architecture for agentic software engineering systems, contrasts the traditional SDLC with an Agentic SDLC in which an orchestrator coordinates specialized sub-agents under human supervision, and compiles evidence on benchmark performance, productivity and labor-market effects. It argues that the field’s central object has shifted from code generation to delegated execution under human supervision, and names five open problems: evaluation beyond SWE-bench, governance, technical debt, skill redistribution and the economics of attention.