Files
openworker/docs/harness/guides/alignment.md
Rohit PrasadandDevika Verma d53e058eda Introduce OpenWorker Harness SDK (in beta), agent service cluster and relevant developer documentation (#844)
Allow developers to easily build AI apps without worrying about the harness part.
Bundles the Rust agent runtime as a SDK (Rust and TypeScript) for building desktop and web applications. Applications can manage a local harness or connect to an authenticated harness service on k8s with separately scaled workers and sandbox execution.

The implementation is primarily contained in harness/, with deployment tooling in deploy/, Markdown guides in docs/harness/guides/, and the static documentation site in docs-site/. It does not replace the existing Python harness or migrate OpenWorker Desktop.

Co-authored-by: Devika Verma <devikaverma11@gmail.com>
2026-11-06 09:45:20 -08:00

1.9 KiB

Agent alignment

An action can be technically permitted and still be unrelated to the user's task. OpenWorker separates what an agent can reach from whether an action is justified by the user's request.

For example, a coding agent may be allowed to read a project file. That does not mean a request to fix a UI bug authorizes uploading the file to an external service.

Choose an approval mode

Mode What happens
Ask Required decisions go to a human; responses and session grants are durable.
Auto-approve A separate reviewer evaluates eligible actions. Uncertainty or reviewer failure pauses for a human.
Bypass Human approval is skipped. Required access policy, sandbox limits and guardrails still apply.

Configure the reviewer's model separately from the working agent. Keep task context, tool effects and source-labelled information available to the review process. Tool output, memory and messages from another agent are data, not new human consent.

Keep the layers independent

  • Access policy and OpenShell constrain resources and execution.
  • Approval/review decides whether an action is appropriate for the task.
  • Guardrails and hooks apply required application rules to content and actions.
  • Compliance records retain the decisions and committed activity for inspection.

Tool categories describe effects; they are not a safety ranking. Calling a tool “read” or “execute” cannot remove another restriction. “Allow always” is scoped to the current session and its bindings, and does not disable mandatory checks. Auto-skipping a question also does not count as approval.

These mechanisms help enforce intent; a model reviewer is not a guarantee of perfect judgment. Evaluate it against representative tasks and adversarial inputs from your application. See guardrails and sessions for configuration and interruption handling.