AI Agents
10 min
Overview (agents are “goal-oriented workflows”)
Kubegrade AI Agents are goal-oriented operational workflows designed to help teams complete specific Kubernetes tasks with strong context and human control.
Built-in agents
Assistant Agent
Purpose:
- Serve as an intelligent high level cluster assistant for users
- Provide explanations of cluster metrics, components, and operational states
- Answer technical questions related to configurations, upgrades, and troubleshooting processes
- Assist users by contextualizing issues, linking relevant documentation, and suggesting next actions
- Support other agents’ workflows by offering guidance and clarification during multi-agent operations
Troubleshooting Agent
Purpose:
- Investigate symptoms across multiple signal layers
- Identify probable causes
- Suggest and optionally prepare remediations
Upgrade Agent
Purpose:
- Analyze target version readiness
- Run preflight checks
- Propose upgrade sequencing and remediation steps
- Create executable and specific upgrade plans
- Generate data necessary to generate PRs for required config changes
PR Generation Agent
Purpose:
- Convert approved remediation plans into repo-specific pull requests
- Preserve format/structure conventions
- Attach rationale and provenance metadata
Custom agents
Custom agents let teams define workflows that combine Kubegrade context with external tools via MCP Servers.
Typical custom agent inputs
- Cluster metadata and object state
- Dependency graph context
- IaC source context (Terraform/Helm/Kustomize)
- Git/GitOps state
- External tools via connectors/MCPs (e.g., Terraform/ArgoCD/etc.)
Example custom use cases
- Policy enforcement workflows
- Scheduled drift checks + PR creation
- Environment readiness checks before releases
- Cost/risk posture reviews across fleet
Prompting patterns + guardrails
Recommended prompting patterns
- Goal + scope + constraints (cluster/env/namespace)
- Desired action mode (analyze-only / propose / PR-ready)
- Risk tolerance and change restrictions
- Output format requirements (summary, checklist, PR body)
Guardrails
- Read-only by default for new workflows
- Policy checks before proposing or executing
- Explicit approval for PR generation/execution
- Scope limits (namespace/workspace/project)
- Logging/audit of agent actions and outputs
Approval + execution controls
Kubegrade keeps AI-assisted workflows under operator control.
Controls to document
- Approval requirements by severity/environment
- Who can run agents vs who can approve results
- Suggest-only mode vs PR generation mode
- Execution via GitOps only (recommended default for prod)
- Time-window restrictions
Feedback loop
Kubegrade can improve agent usefulness over time through operator feedback.
Feedback signals (recommended)
- Accepted vs rejected suggestions
- Edits made before PR generation
- Rollback frequency
- Time to resolution after applying recommendation
- “Incorrect root cause” or “missing context” tags
Why it matters
- Better remediation quality
- Less noisy suggestions
- More consistent outputs across teams