kontext

Agents are workloads

Kontext is a Kubernetes control plane for AI agents. An Agent is a custom resource reconciled into real Pods: apply it, watch runtime logs, read the result. No new platform to learn.

apiVersion: kontext.dev/v1alpha1
kind: AgentRun
metadata:
  name: review
spec:
  goal: "Summarize the failure modes in this release"
  provider: echo
  model: echo-model
  runtime:
    image: ghcr.io/mfs-code/kontext-echo:v0.1.0-alpha.2

$ kubectl apply -f run.yaml
$ kubectl logs -f run-review
$ kubectl get agentrun review \
    -o jsonpath='{.status.result}'

alpha · install without cloning · docs

Why Kubernetes

Scheduling, secrets, RBAC, budgets, logs, restarts: agents need what Kubernetes already does. Kontext adds two resources, Agent and AgentRun, and leaves the rest to the cluster. Task agents are reusable one-shot templates. Scheduled agents mint one-shot runs from cron slots. Service agents are re-cast automatically when they die. Wallclock budgets are enforced by the controller. Results live in status, queryable like any other object.

Bring your own runtime

Any container can be an agent. Ship your own image, or start with the echo and reference runtimes that come with a release.