How well do AI agents diagnose a broken Kubernetes cluster?
We break real clusters and measure what agents do about it. Three studies, each answering one question: does what the agent is given matter, which MCP server is fastest, and which finished product diagnoses correctly. We build Radar, so every number comes with the dataset behind it.
With Radar's MCP server, the same agent diagnosed faults 3× faster on average than with raw kubectl, with slightly higher accuracy. Radar was 2.3× faster on average than the next-fastest Kubernetes MCP server. Against 5 AI SRE tools, including HolmesGPT, kagent and AWS DevOps Agent, Claude Code on Radar's MCP was the most accurate, the fastest and the cheapest.
- One open benchmark suite
- SREGym, from Microsoft and UIUC
- One live cluster
- 3-node Amazon EKS, real applications, injected faults
- One judge
- Claude Opus 5 on SREGym's own rubric, identical for every arm
- Every dataset downloadable
- per-fault scores and times as JSON, plus costs for the AI SRE study
- Compare results within one study
- each study ran all its arms on the same faults at the same time. Between studies the faults, date and cluster load differ, so compare the gaps within a study
0150 faults · same agent, kubectl vs Radar
Does an MCP server beat raw kubectl?
Same agent, same model, same faults - the only thing that changed was what we handed it to work with.
Average time to a diagnosis, all 50 faultslower is better
- Radar MCP64s
- Raw kubectl191s
3× faster on average · slightly higher accuracy
Replay any scenario call-by-call0225 faults · 5 MCP servers and kubectl
Which Kubernetes MCP server is fastest?
If what the agent is given matters, which MCP server do you install? Four of them and raw kubectl, one agent driving all five.
Average time to a diagnosis, all 25 faultslower is better
- Radar57s
- Flux159/mcp-server-kubernetes135s
- containers/kubernetes-mcp-server149s
- k8sgpt174s
- kubectl249s
- Amazon EKS MCP321s
2.3× faster on average than the next-best Kubernetes MCP server
See the ranking0350 faults · Radar MCP and 5 AI SRE tools
Which AI SRE tool diagnoses correctly?
Claude Code on Radar's MCP against five AI SRE tools, each with its own agent and prompt, including AWS's managed DevOps Agent.
Accuracy against average answer time, 50 faultsbest is top-left · correct and cost per fault on each labelsorted by correct answers
| Product | Correct | Average |
|---|---|---|
| Radar MCP44/50 · $0.27 | 88% | 79s |
| kubectl-ai41/50 · $1.67 | 82% | 288s |
| HolmesGPT38/50 · $0.37 | 76% | 176s |
| OpenSRE38/50 · $2.07 | 76% | 151s |
| kagent38/50 · $2.26 | 76% | 181s |
| AWS DevOps Agent33/50 · $1.27 | 66% | 123s |
Radar driven by Claude Code. All tools used Claude Sonnet 5 via Amazon Bedrock, except AWS DevOps Agent, which runs its own proprietary model and harness.
See all 50 faultsRun Radar on your own cluster
Radar is open source, one binary. Run it on your laptop or in your cluster.
$curl -fsSL https://get.radarhq.io | sh && kubectl radarApache 2.0 · No account for local use · Run Radar OSS forever