Agent compute
Compute is the infrastructure that runs an Agent. The installation operator selects a Compute Driver; an Agent owner chooses an execution mode supported by that Driver. You cannot select a different Compute Driver for an individual Agent. If you are setting up your own first installation, use the Kubernetes quickstart.
Choose an execution mode
| Mode and Harness | Deployment and use |
|---|---|
| Embedded OpenClaw | One gateway runs the OpenClaw Harness. Use it on Kubernetes or an operator-managed SSH host. |
| Dedicated Codex | A gateway connects to a separate Codex Harness on Kubernetes. Use it for Codex, a ChatGPT service account, or a supported external channel. |
| Dedicated native OpenClaw (experimental) | A gateway connects to a native worker. Requires a provisioning SandboxDriver with all three isolation facets. The bundled OpenShell path has upstream blockers and is verification-only. |
The mode belongs to the Agent. Its Configuration selects the model and Harness, and the two must agree. The HTTP API defaults a new Agent to embedded when the mode is omitted; the first-Agent guide uses embedded OpenClaw on Kubernetes. See Harness for the supported model and credential combinations.
What the bundled Drivers support
| Driver | Current use |
|---|---|
| Kubernetes | Embedded OpenClaw and dedicated Codex. Dedicated native OpenClaw additionally requires a full-facet provisioning SandboxDriver; the bundled OpenShell path is verification-only until its upstream projection requirements are met. OpenClaw requires a managed OpenAI API key; dedicated Codex also accepts an issued ChatGPT service account credential. |
| SSH | Embedded OpenClaw on operator-managed Linux hosts. The operator supplies the model credential on the host; the control plane does not check model access. Dedicated Codex and platform-managed model credentials are unavailable. |
| Docker | Local control-plane development only with the current Agent authentication contract. Its default Docker/Podman profile cannot admit a new Agent deployment. |
A Sandbox changes how a supported workload is contained; it does not replace Compute or choose an Agent's model. It is optional for other topologies and mandatory for dedicated native OpenClaw.
Deployment and availability
A deployment creates a new Agent Revision. The control plane queues work and can select the new revision while activation is still in progress. Neither the accepted request nor the console's selected revision shows live health. Use Troubleshoot for status checks. For a real model response, follow the local Kubernetes first-Agent guide or the production Kubernetes verification guide. For SSH deployments, use the SSH Driver's credential and verification guidance.
On Kubernetes, replacing an embedded Agent can interrupt service before the new model credential is accepted. If the replacement cannot start, there is no automatic rollback. See Harness isolation and activation before using that mode for an availability-sensitive Agent.