AgentMoE-Test: A Mixture-of-Experts Agentic Framework with Graph Retrieval-Augmented Generation for Autonomous Sprint Automation in Playwright CI/CD Pipelines

Abstract

Modern continuous integration and continuous delivery (CI/CD) pipelines require reliable end-to-end testing, yet Playwright suites remain costly to create, maintain, and repair. Existing large-language-model agents often use incomplete project context, repeat failed actions, and generate brittle tests after interface or sprint changes. This paper proposes AgentMoE-Test, a mixture-of-experts framework combining graph retrieval-augmented generation with autonomous sprint automation. Sparse routing assigns sprint tasks to specialized experts for requirement analysis, test planning, Playwright code generation, execution, diagnosis, and bounded repair. A project knowledge graph $G=(V, E)$ connects requirements, code, interface elements, tests, traces, and defects. Task-specific retrieval supplies grounded evidence and preserves traceability. Simulation experiments against SWE-agent, Agentless, CoverUp, and TestART produced 91.6% requirement coverage, 92.8% code coverage, 94.2% generation accuracy, 93.7% pass rate, 91.4% defect detection, and 88.9% repair success. AgentMoE-Test recorded 3.1% flakiness, 19.6 minutes of pipeline duration, and $2.75 token cost per task. Relative to the strongest baseline, repair success increased by 27.36%, while flakiness, duration, and cost decreased by 56.94 %, 29.50 %, and 11.29 %. The results support reliable, efficient, and auditable sprint-to-test automation.