How to triage Kubernetes incidents with Unknown
In a broken-image test on k3s, NeuroScale diagnosed the issue but withheld remediation because runbook similarity was below its gate, escalating with a rollback command.
dev.toPrompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Triage [Kubernetes incident] using [cluster event data] and [runbook library]. Analyze the cause, retrieve a matching runbook, and prepare a rollback command. Execute only if [confidence requirement], [similarity threshold], and [risk requirement] all pass; otherwise request human approval.
Reach people exploring AI agents
See placements →What happened
The broken-image incident was correctly diagnosed, but a 0.59 runbook score fell below the 0.65 gate, so the system escalated with a rollback command instead of executing.
Reported by sodiqjimoh on dev.toHow the agent did it
- Deploy NeuroScale on an ECS-hosted k3s cluster.
- Push a broken image tag to a sample checkout service.
- Diagnose the incident with Qwen-Max.
Where a person stepped in
The tested incident remained open awaiting human approval or rejection because its runbook retrieval score failed the configured threshold.
A person approved each step before the agent continued.
NeuroScale Autopilot is a five-agent Kubernetes incident-response system built for a hackathon. Its pipeline detects cluster events, analyzes root causes with Qwen-Max, retrieves runbooks through embeddings and FAISS, executes approved remediation, and uses Qwen-Turbo to summarize escalations.
The author deployed it on an Alibaba Cloud ECS instance running k3s, then deliberately pushed a broken image tag to a sample checkout service. The analyzer correctly diagnosed the issue and rated its confidence high and risk low. However, the retrieved runbook scored below the configured similarity threshold. The system held the incident open instead of executing the proposed fix, attached a rollback command, and waited for human approval.
Automatic execution requires all three conditions: high analysis confidence, a runbook similarity score clearing the threshold, and low risk. Failure of any condition routes the incident to human review. The author also reports that failed Qwen calls during infrastructure testing always caused escalation rather than an attempted guess.
The observed result demonstrates a withheld action, not a completed automated repair. The deployment produced human-ready decisions and attached rollback commands to remediation plans. Who it's for: platform engineers evaluating incident agents that separate diagnosis from permission to act and explicitly expose uncertain runbook matches.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: dev.to · sodiqjimoh
Frequently asked questions
Why was the tested incident escalated?
Its runbook similarity score was 0.59, below the 0.65 threshold, despite high analysis confidence and low assessed risk.
What happened when Qwen calls failed?
The author reports that failures during infrastructure testing always escalated to a human rather than prompting the system to guess.
More details
Remaining steps
- Retrieve a runbook and evaluate its similarity score.
- Hold remediation and escalate with a prepared rollback command.
About this account
The system detects and analyzes incidents and retrieves remediation plans; execution is gated on confidence, runbook similarity, and assessed risk.
Platform: dev.to. Attribution: https://dev.to/sodiqjimoh/the-agent-that-knows-when-not-to-act-building-neuroscale-autopilot-on-qwen-cloud-1lld. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: sodiqjimoh on dev.to
Source published 7 Jul 2026
Source checked 14 Oct 2026
Tools and topics
NeuroScale Autopilot · Qwen-Max · Qwen-Turbo · text-embedding-v3 · FAISS · Alibaba Cloud ECS · k3s
Kubernetes · Incident response
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.