This dissertation investigates whether LLM-based agents can autonomously implement, test, and validate software within an Agile DevOps workflow while maintaining effective human oversight. It presents AgileAI, a multi-agent system that integrates with GitLab's issue tracker and continuous integration / continuous deployment (CI/CD) pipeline. A human Product Owner approves user stories via labels; a planner agent groups approved stories into sprints; developer and test-generation agents implement each sprint; and a CI repair loop iteratively patches code in response to pipeline failures. The reviewer retains authority through three checkpoints: story approval, review of the deployed application, and merge request approval. Three controlled experiments compared four pipeline configurations against a pre-registered acceptance checklist on a Flask notes application, evaluated across two story batches modelling an initial delivery and a follow-on extension. The four configurations were: a sprint-informed serial pipeline with CI repair, a one-shot waterfall baseline, a sprint pipeline with the repair loop disabled, and a parallel pipeline. Sprint-informed achieved near-perfect, consistent results (mean 33.6 of 35 requirements; median 33; range 33-35) across all five runs. Waterfall was faster but highly variable, with three of five runs producing CI-green applications that failed entirely at deployment. Sprint-no-CI exhibited a 100% build failure rate, confirming the repair loop is necessary for deployable output. Parallel execution showed no meaningful throughput gain, with three of five runs collapsing before reaching CI green. The findings support a narrower but more useful claim than full autonomy: agents reliably perform implementation, testing, and repair when humans retain orchestration and final-verification roles.