Agentic AI in Software Engineering: Opportunities, Risks, and a Human-Controlled Development Framework

Abstract

Agentic artificial intelligence is moving software engineering beyond prompt-based assistance toward systems that can plan tasks, invoke tools, modify repositories, execute tests, inspect results, and iterate with limited human intervention. This transition changes the engineering problem: the central question is no longer only whether an AI model can generate useful code, but whether an AI agent can act safely and reliably inside a software delivery system. This paper presents a structured review of agentic AI for software engineering and proposes a Human-Controlled Agentic Development Framework (HC-ADF). The review synthesizes foundational work on reasoning-and-acting systems, tool use, autonomous agents, multi-agent orchestration, coding agents, repository-level benchmarks, developer productivity, security, and AI governance. The proposed framework organizes agentic software development around six control stages: intent definition, planning, bounded execution, evidence-based verification, human authorization, and controlled release. It further introduces an autonomy ladder, risk-based approval gates, identity and tool permissions, sandboxing, traceability requirements, and a multidimensional evaluation model covering task success, code quality, security, change scope, verification effort, delivery stability, and human override behavior. The paper treats agentic AI as a software-delivery participant operating under explicit constraints rather than as an unrestricted autonomous programmer. Recent evidence from SWE-bench, SWE-Lancer, DORA, NIST, and OWASP shows substantial capability progress alongside persistent limitations in long-horizon reasoning, context handling, verification, security, and organizational control.