Type: url
Original: https://arxiv.org/pdf/2605.20456
Status: generated
Summary
A synthesis and process proposal arguing that the central problem in agentic coding is no longer prompt engineering but engineering process control. Reviewing evidence from productivity trials, GitHub-scale adoption studies, repository-configuration work, issue-resolution benchmarks and hardware verification research, the author holds that agentic AI can generate plausible engineering artifacts faster than humans can inspect them, so the bottleneck shifts to specification quality, execution context, verification and traceability. The paper proposes Agentic Agile-V, pairing an Agile-V lifecycle backbone with a task-level SCOPE-V loop, and sets out a conversation-to-contract gate, a minimum input-artifact taxonomy, risk-adaptive acceptance gates and an evidence-bundle model for accepting agent-generated work.