Agentic AI in Software Delivery: Autonomous Pull Request Generation, Testing, and Deployment in Production DevOps Pipelines

Abstract

Agentic AI systems�autonomous LLM-based agents capable of multi-step tool use�are beginning to participate directly in software delivery pipelines. This paper presents a controlled production study of an agentic AI system that autonomously generates pull requests for dependency upgrades and bug fixes, executes test suites, interprets failures, and initiates deployment workflows under human oversight gates. Deployed across 4 organisations over 8 months, the agent successfully resolved 49% of assigned tasks end-to-end without human intervention, reducing toil by an estimated 1,100 engineer-hours per quarter. Safety boundaries, escalation protocols, and auditability requirements for production agentic deployment are formalised. Failure mode analysis reveals that ambiguous acceptance criteria and flaky tests are the dominant blockers to autonomous resolution.