Agentic AI for Autonomous Software Development

This paper surveys agentic AI for autonomous software development, covering a six-stage lifecycle, a five-level autonomy scale, comparative agent architectures, and benchmarks like HumanEval, MBPP, SWE-bench, and InterCode. It argues that benchmark scores do not equal dependable production performance, and proposes a risk-aware reference architecture combining sandboxing, least-privilege tools, test-driven iteration, provenance records, and human approval gates, along with open challenges in planning, verification, evaluation, cost, security, and human-agent coordination.