Kubernetes Jobs
Heavy agent work becomes schedulable Jobs with isolation, placement, and concurrency controls.
Agents as cluster jobs, not chatbots
Durable multi-harness agent loops on Kubernetes. Turn builds, suites, scans, indexes, and long research into isolated Jobs with composable profiles, durable storage, and an observable run record.
A chatbot answers a tenant question in seconds. Platform work — multi-file changes, test suites, security analysis, recurring maintenance — needs identity, capacity, isolation, and a result that remains useful after the conversation ends.
Heavy agent work becomes schedulable Jobs with isolation, placement, and concurrency controls.
Codex, OpenCode, Hermes, OpenClaw, Grok Build, Pi, or a custom harness behind one run contract.
Change a skill, tool, harness, profile, or schedule independently — keep change local.
No Anvil Primaris or control-plane lock-in required. Apache-2.0 operator you can run anywhere.
The important design choice is not the model provider. It is the set of objects that describe what the work is, how it runs, and when it should start. Read the essay.
Append-only execution record. One run, one Job, one durable result. A retry is a new run.
Reusable role: scope, policy, standing prompt, and composition for a worker without embedding pod details.
Execution envelope: backend, image, identity, credentials, storage, limits, placement, timeout.
Instructions stay separate from executables. Skills teach when; tools install and prove clients.
Cadence and launch policy for append-only runs — suspension, concurrency, named templates, backoff.
Optional read API and live SSE logs so operators can inspect runs with their own OIDC-protected deployment.
Full operator docs live on this site — getting started, harnesses, security, streams, and more — with diagrams that match how the system actually runs.