NeuroScale Autopilot Uses a Trust Layer to Decide When AI Should Not Act

A developer built NeuroScale Autopilot, a five-agent Kubernetes incident-response system, for the Qwen Cloud Global AI Hackathon, deliberately prioritizing caution over speed. The system uses three AI models from the Qwen family — Qwen-Max, text-embedding-v3, and Qwen-Turbo — assigned to distinct roles: root cause analysis, runbook retrieval, and human-readable escalation respectively. At the core of the design is a Trust Layer gate that requires high analyzer confidence, a runbook similarity score above 0.65, and a low risk rating before any automated action is taken. During a live test on an Alibaba Cloud ECS instance running a k3s cluster, the system correctly diagnosed a broken image tag but withheld the fix because the runbook match scored 0.59, just below the threshold. Rather than proceeding on a near-match, the agent escalated to a human with the rollback command already prepared, illustrating the project's central argument that confidence alone is not sufficient grounds for autonomous action.