Mabl's engineering team built an auto-reliability pipeline that automatically investigates, fixes, and tests production failures in their AI test authoring system. The pipeline chains a Session Review Agent (SRA) — which has no code access and excels at symptom detection — with Claude-powered investigator and fix-worker agents that verify root causes against the actual codebase before writing any code. Key design decisions include an adversarial investigation phase that debunks incorrect SRA hypotheses, a six-lens simulated peer review panel, and a pre-merge headless execution step. Hard-won lessons include the need for explicit liveness rules to prevent agents from idling, convergence policies to stop infinite refinement loops, and isolating test suite execution to CI rather than the orchestration host. The system reduced human touchpoints to three: batch kickoff, architectural policy calls, and final PR merge.
