Beyond Static Gates: Closing the Detect–Fix–Learn Loop in Agentic CI/CD Quality Assurance

Static quality gates in continuous integration pipelines suffer from three compounding limitations: single-model risk assessment provides no uncertainty signal, fixed escalation thresholds accumulate miscalibration over time, and detected defects require manual developer remediation. This paper extends the K11tech Agentic AI QA System [1] with three interlocking innovations that close a detect-fix-learn feedback loop. First, a Multi-LLM Consensus Gate queries GPT-4o, Claude Sonnet, and Gemini 1.5 Pro in parallel and forces human-in-the-loop (HITL) escalation when models disagree-treating epistemic uncertainty as a first-class safety signal. Second, an AutoRemediationAgent generates confidence-gated unified-diff patches for five safe defect classes and opens remediation pull requests via the GitHub MCP server, eliminating post-detection developer toil for deterministic defects. Third, an Adaptive HITL Threshold Learner applies an exponential moving average over reviewer approval rates, continuously recalibrating the escalation threshold within safety bounds. Simulation over 500 synthetic reviewer decisions demonstrates threshold convergence within 80 decisions (MAE = 0.018) and a 34% reduction in unnecessary HITL activations vs the fixed-threshold baseline, with zero increase in production escape rate. The three mechanisms compose safely: the consensus gate provides the uncertainty signal the learner uses to weight its updates.