Human Verification in Agentic Software Engineering: A Mixed-Methods Study of Risk-Scoped Autonomy

AI coding assistants have moved from novelty to infrastructure in software engineering, increasingly acting as agents that plan and execute multi-step development tasks on their own. We study this shift through a mixed-methods design, combining a survey of 200 Singapore-based developers and non-developers with 44 interviews of engineering stakeholders worldwide. We find that AI coding tools have were widely used in our survey sample (86%); among tool users, 70% reported higher personal productivity. However, agentic autonomy, governance, and review practices have not scaled at the same pace. No agentic-AI user reported granting an agent full autonomy to commit or deploy, and governance guidelines remain informal for 40 of organizations. Reviewing and validating AI-generated code is emerging as the most important future skill, ahead of traditional programming ability, and our results point to risk-scoped autonomy and outcome measurement as the two levers most likely to close the gap between how fast organizations adopt AI and how well they govern it.