Agents

An Agent is an AI workload that you create and deploy inside a Namespace. It has its own identity and, once deployed, its own gateway. Its Harness calls the model and runs tools. If you are new to OpenClaw Enterprise, set up an OpenShell Codex Agent and verify that it can answer.

How an Agent runs

1. Create a draft. Give the Agent a name, select an execution mode, and choose a Configuration and model credential. Creation saves the draft; it does not start anything.
2. Deploy the draft. The control plane saves an Agent revision for the new version, then queues work to start it. Editing a draft later does not change the running Agent; deploy again to apply the new settings.
3. Check the result. A selected version means deployment progressed. The console's Logs tab shows Pod status, restarts, Events and container output. On Kubernetes, Run diagnostics for this version checks only the Slack channel. Neither tests the model, and OCE has no browser chat. Check an actual model response before treating a new deployment as working.
4. Stop when needed. Stopping ends execution and routing but preserves version history, credentials, and persistent state. Deploying again starts a new version. To remove the Agent and its versions, delete it. There is no rollback endpoint.

Work with an Agent

• Compute: choose embedded OpenClaw or dedicated Codex and see which infrastructures support them.
• Agent revisions: update configuration, deploy, and inspect what changed.
• Workspace files: edit the running Agent's instructions and identity files.
• Plugins: add selected tools and decide which actions require approval.
• Troubleshoot: investigate deployment, model access, and uncertain requests.

For exact permissions, API behavior, and credential rules, use the Agent reference.