Kontext runs AI agents as Kubernetes workloads with Agent and AgentRun resources.
Introduction
Kontext is a Kubernetes-native control plane for running, governing, and observing AI agents as production workloads.
The thesis: agents are workloads. An agent should not live in a screen
session or behind a bespoke orchestration service. It should be a resource your
cluster understands — created with kubectl apply, observed with
kubectl logs -f, governed by RBAC and budgets, and restarted by a controller
when it dies.
Two resources
| Kontext | Kubernetes analogue | Behavior |
|---|---|---|
Agent (mode Service) | Deployment | Always-on. The controller keeps one live AgentRun and re-casts it when it exits. |
Agent (mode Task) | reusable template | Creating the Agent runs nothing. Sparse named invocations become immutable owned AgentRuns. |
Agent (mode Scheduled) | CronJob | Mints one-shot AgentRuns from standard five-field cron slots with safe overlap and deadline policy. |
AgentRun | Job / Pod | One bounded execution. Owns one Pod or is warm-delivered to a standing Service runtime, and holds .status.result. |
You can also create a standalone AgentRun without an owning Agent — useful
for ad-hoc dispatch and demos.
Next steps
- Install without cloning
- Learn the resource model
- Invoke a reusable Task workload
- Run a Scheduled workload
- Keep a Service workload available
- Read the API spec