Skip to main content

Safe Execution Model

Overview​

OpsWorker is designed around a fundamental principle: AI investigates, humans decide. OpsWorker never modifies your Kubernetes cluster state: it recommends actions for engineers to review and execute. The one exception is source control, where OpsWorker can open a pull or merge request, which is always human-reviewed before it lands (see below).

Read-Only Kubernetes Access​

For your Kubernetes cluster, OpsWorker performs only read operations:

Allowed OperationsBlocked Operations
kubectl getkubectl apply
kubectl describekubectl delete
kubectl logskubectl patch
kubectl top (metrics)kubectl exec
List eventskubectl scale
Check endpointskubectl rollout

Against the cluster, the agent cannot create, update, delete, scale, roll out, or exec into containers.

The One Write Path: Git PR/MR​

The only change OpsWorker can initiate is in source control, and only via pull/merge requests:

  • The Source Code & Repository Agent can create a branch and open a PR (GitHub) or MR (GitLab) with a proposed fix.
  • The Resource Optimizer can open a PR/MR with a right-sized manifest after you confirm (a draft if confidence is not high).
  • Nothing is merged automatically as a background job. A PR/MR exists only when a tool call returned a live URL during the session, and a human reviews and approves it before it merges.

Recommendations, Not Actions​

Every investigation produces recommendations, specific steps with kubectl commands. These are:

  • Displayed in Slack and the portal for human review
  • Never auto-executed on your cluster
  • Copy-paste ready for engineers who decide to apply them
  • Contextualized with the reasoning behind each recommendation

Why This Model​

  • Safety: No risk of AI-initiated changes in production
  • Compliance: Humans maintain full control of cluster modifications
  • Trust: Engineers can verify recommendations before acting
  • Audit trail: Every investigation and its recommendations are logged

Human-in-the-Loop​

The workflow ensures human oversight at the action stage:

flowchart LR
    A[Alert] --> B[AI Investigation]
    B --> C[Recommendations generated]
    C --> D[Engineer reviews]
    D --> E{Approve?}
    E -->|Yes| F[Engineer executes commands]
    E -->|No| G[Engineer takes alternative action]

OpsWorker handles the investigation (the time-consuming, repetitive part). Humans handle the decision-making and execution (the part that requires judgment and accountability).

Next Steps​