Agents need relationships and changes, not only YAML
Radar exposes Kubernetes relationships, health, deduplicated events, change diffs and filtered logs to both its UI and MCP clients. Raw kubectl investigations spend calls reconstructing this context. Informer-based observation lets Radar supply existing relationships and recent changes so an agent can follow dependencies from an unhealthy resource.
For a stuck rollout, the UI and agent can inspect the same Pods, controllers and events before requesting logs. This reduces context assembly but cannot substitute for absent application telemetry. A connected model receives what its agent reads; local tooling does not imply that a remote model sees no cluster data.
The boundary of the read-only endpoint
/mcp-readonly excludes write tools. The full /mcp endpoint offers restart, scale and apply operations, and some diagnostics can create temporary probe Pods. MCP read-only settings do not disable an agent’s shell, other MCP servers or all ordinary UI writes.
Investigation and remediation therefore have distinct authority. Secret fields are withheld and known patterns scrubbed, but pattern matching cannot guarantee removal of all sensitive data. Deployment decisions depend on the agent’s execution mode and Kubernetes credentials, not the endpoint name alone.
Where the threefold result comes from
The authors inject SREGym faults into a three-node EKS cluster. The first study holds agent and model fixed across 50 faults, changing Radar MCP versus raw kubectl; mean time to a submitted diagnosis is 64 versus 191 seconds. Wrong answers also count in that timing, so it is not mean successful repair time. Separate MCP and full-product studies use different faults, dates and cluster loads and cannot be ranked across tables.
My interpretation is that the comparison motivates studying structured context rather than only stronger models. It is author-run and uses an LLM judge. Real incidents introduce different fault distributions, histories and missing observations. Faster answers remain distinct from correctly diagnosed and authorized service recovery.
Source: Original engineering article