​
Completed state

At the end of this guide, one Agent has a saved runtime configuration, its environment test has an understood result, and a low-risk Agent Run has left a visible result or blocker with transcript or raw output evidence. The Agent, Agent Run, and runtime remain separate. The Agent owns the role, the Run records one attempt, and the runtime performs that attempt. Compare operator-facing choices in Runtime types.

​
Prerequisites

You need permission to edit an Agent and the runtime on the same host or network where Rudder will invoke it. For a local runtime, install and authenticate the matching CLI first. For an external service, prepare its reachable endpoint and credentials. Keep secrets in the host environment or Rudder secret fields, not in committed configuration. This walkthrough uses an illustrative release writer with Codex local. The same verification pattern applies to other operator-facing runtimes, but their capabilities differ.

​
Configure the release writer

  1. Open the release writer’s Agent detail and edit its Runtime configuration.
  2. Choose Codex (local). If the default CLI is missing, Rudder shows setup guidance but still permits a supported custom command.
  3. Choose the model and thinking effort offered for that model. Changing model clears an effort value that the new model does not support.
  4. Add a role-specific instructions file only if the Agent needs one. Use an absolute path that the Rudder host can read.
  5. Review broad options such as web search or sandbox bypass. Enable them only when this trusted local work requires them.
  6. Run Test runtime chain. Treat pass as ready, warn as visible guidance that must be understood, and fail as a failed probe that needs repair.
  7. Save the Agent, then ask it in Chat to draft a two-line release headline.
  8. Open the resulting Agent Run and confirm the selected runtime, result, transcript or raw output, and any usage or cost evidence.
Rudder supplies first-party control tools automatically to supported local runtimes. Do not add a second custom Rudder MCP server. Browser tools appear only for an eligible local run while the Built-in Browser setting is enabled.

​
Success signal

The environment result is either pass or a warn whose limits you understand, and the real run starts under the selected runtime. The Chat shows a finished answer or explicit blocker, while Agent Run detail preserves evidence underneath it. A successful probe without a real run is not enough proof.

​
Recovery

If the probe reports command not found, confirm the CLI is installed on the Rudder host and available on its PATH, or set a supported custom command. If authentication fails, log in through that runtime outside Rudder and test again. For an unreadable instructions file or workspace, correct the absolute path and host permissions before retrying. If a warn becomes a failed real run, use the Run’s terminal error and raw evidence to repair the exact provider, model, endpoint, or permission. Do not switch the Agent to active recurring work until the low-risk run succeeds.