kagent Assess

platformagentsopen-sourcekubernetesoperationsassessdevmay-2026kagent

Overview

kagent connects Kubernetes operations with AI agents for diagnostics, remediation suggestions, and Day-2 automation (kagent).

Assess with read-only RBAC first. Keep auto-remediation out of production clusters until actions are idempotent, approval-gated, and audited.

Adoption Signals

  • CNCF status: kagent was accepted into the CNCF on 22 May 2025 at Sandbox maturity, about two months after Solo.io announced it in March 2025 (CNCF project page, 100 days of kagent).
  • Contributor base: in its first 100 days kagent reached more than 100 contributors, over 85% of them from outside Solo.io; by September 2026 the repository had about 3,900 stars and 800 forks (100 days of kagent, kagent GitHub).
  • Active development toward 1.0: v0.10.2 is the current stable release (23 September 2026), and v1.0.0 alphas add sandbox runtimes and explicit agent sessions (kagent releases).
  • Commercial backing: Solo.io offers Solo Enterprise for kagent with global SLAs and architecture guidance from the maintainers, and pairs kagent with agentregistry for cataloguing agents, MCP tools, and skills (kagent enterprise, Solo.io AI agents).
  • Safety primitives: agents can list tools under requireApproval so that apply, patch, and delete calls pause for a human decision while read-only calls run freely (kagent human-in-the-loop).

Risks

  • Agent with cluster credentials. An agent that can call k8s_apply_manifest or k8s_delete_resource under a broad service account turns a prompt injection in logs, events, or issue text into a cluster change. Approval is opt-in: "Tools run freely by default" (kagent human-in-the-loop).
  • Maturity. The CNCF describes Sandbox projects as "experimental projects not yet widely tested in production", the CNCF project page rates kagent's health as "Fair (67)", and the API is still changing on the way to 1.0 (CNCF Sandbox projects, CNCF project page, kagent releases).
  • Thin production evidence. Public material consists of community use cases, demos, and vendor pages rather than named production deployments with outcomes (100 days of kagent).
  • Model data flow. Cluster state, logs, and secrets referenced in tool output are sent to the configured LLM provider unless a self-hosted model is used, which matters for regulated workloads.

Pros & Cons

Advantages

  • Kubernetes-native design: agents, tools, and model configs are custom resources managed with kubectl, GitOps, and RBAC, so platform teams can review and roll back agent definitions like any other manifest.
  • Ships MCP tool servers for Kubernetes, Istio, Helm, Argo, and Prometheus, supports OpenAI, Azure OpenAI, Anthropic, Vertex AI, and Ollama, and emits OpenTelemetry traces for every agent run.
  • Neutral home and broad contributor base: a CNCF Sandbox project since May 2025 under Apache-2.0, with more than 85% of its first 100 contributors from outside Solo.io, the original creator.

Disadvantages

  • Still pre-1.0 (v0.10.2 stable, v1.0.0 in alpha as of September 2026), and Sandbox is CNCF's "experimental, not yet widely tested in production" maturity level.
  • Tools run without approval by default; human-in-the-loop gating of destructive calls such as k8s_delete_resource or k8s_apply_manifest has to be declared explicitly per agent.
  • No named production adopters or independent case studies were found; the public evidence is conference talks, community demos, and Solo.io's commercial offering.

Recommendation

Assess kagent as a read-only diagnostics assistant on a non-production cluster: bind its service account to view-only RBAC, route model calls to an approved or self-hosted provider, and export its OpenTelemetry traces to the existing observability stack. Measure whether it shortens troubleshooting compared with the runbooks and dashboards already in use.

Only after that, test write actions in a staging cluster with every mutating tool listed under requireApproval, namespaced permissions, and GitOps reconciliation (Argo CD or Flux) as the path for durable changes. Re-evaluate once kagent reaches 1.0 and named production users publish results; until then it is not a candidate for unattended remediation in production.

Sources

Consulting

From radar to implementation

You have assessed the technologies — we help you bring them into production systems safely and measurably.

AI Trainings

AI Trainings

We make your teams AI-ready. Our trainings turn abstract buzzwords into applicable skills. Clear. Practical. Built for doers.

Learn more