Kubernetes Agent Swarm — Platform Operations

A multi-agent system for Kubernetes and OpenShift platform operations. Seven specialized agents work together as a coordinated swarm.

Runtime Requirements

Requirement	Required	Description
INLINECODE0	✅ Yes	Kubernetes CLI — must be in PATH
INLINECODE1

Optional | OpenShift CLI — needed for OCP/ROSA/ARO | | helm | Optional | For GitOps agent Helm operations | | jq | Optional | For JSON output parsing | | KUBECONFIG | ✅ Yes | Cluster access via env var or ~/.kube/config |

Optional cloud CLIs (aws, az, gcloud, rosa) — only needed for managed cluster operations.

Installation

CODEBLOCK0

Or install individual agents:
CODEBLOCK1

The Swarm — Agent Roster

Agent	Code Name	Domain
Orchestrator	Jarvis	Task routing, coordination, standups
Cluster Ops

Atlas | Cluster lifecycle, nodes, upgrades | | GitOps | Flow | ArgoCD, Helm, Kustomize, deploys | | Security | Shield | RBAC, policies, secrets, scanning | | Observability | Pulse | Metrics, logs, alerts, incidents | | Artifacts | Cache | Registries, SBOM, promotion, CVEs | | Developer Experience | Desk | Namespaces, onboarding, support |

How It Works

This is an instruction-only skill. Agents receive markdown instructions describing what commands to run and how to interpret output. No executable scripts are included — the agent translates instructions into actions using the host's installed CLI tools.

Session Setup

Before using the swarm, establish cluster context:

CODEBLOCK2

Agent Communication

Agents communicate via @mentions in shared task comments:
CODEBLOCK3

Escalation Path

1. 1. Agent detects issue
2. Agent attempts resolution within guardrails
3. If blocked → @mention another agent or escalate to human
4. P1 incidents → all relevant agents auto-notified

Heartbeat Schedule

CODEBLOCK4

Agent Capabilities

What Agents CAN Do

• - Read cluster state (kubectl get, kubectl describe, oc get)
• Deploy via GitOps (argocd app sync, Flux reconciliation)
• Create documentation and reports
• Investigate and triage incidents
• Provision standard resources (namespaces, quotas, RBAC)
• Run health checks and audits
• Query metrics and logs

What Agents CANNOT Do (Human-in-the-Loop Required)

• - Delete production resources
• Modify cluster-wide policies
• Make direct changes to secrets without rotation workflow
• Perform irreversible cluster upgrades
• Approve production deployments (can prepare, human approves)

Key Principles

• - Roles over genericism — Each agent has a defined domain
• Files over mental notes — Only files persist between sessions
• Human-in-the-loop — Critical actions require approval
• Guardrails over freedom — Define what agents can and cannot do
• Audit everything — Every action logged

File Structure

CODEBLOCK5

Detailed Agent Documentation

See individual SKILL.md files for each agent's full capabilities, personality, and workflow instructions.