An agent says it finished. You still need to reconstruct what changed and why.
The task, run history, and result are connected for review.
Ask for evidence before the run begins
A good review starts in the brief. Specify the artifact you expect, the checks that matter, and the actions that need human judgment. For a coding task, that could mean changed files, test output, and a branch or pull request.
Norc keeps run conversations and results with the work. Inspect the transcript when a conclusion needs explanation, rather than treating a status label as proof that the outcome is correct.
Intervene while the work is happening
Run visibility lets you see when an agent has misunderstood the task or needs missing context. You can steer or stop an active run instead of waiting for it to consume the full time and budget.
Project autonomy settings control how work proceeds. Define human decision points around the consequences of the task, such as changes to a production environment or a customer-facing deliverable.
Choose what happens when an agent is blocked
A blocked agent does not always need the same response. Configure a stand-in, waiting and retrying, or a human decision according to the failure reason.
A stand-in can work with the original agent’s persona, pinned skills, and assigned secrets in a fresh session. Recovery helps work continue; it does not prove the answer is correct or remove the need to review sensitive actions.
Practical questions.
Can I stop a running agent?
Norc exposes controls to steer or stop an active run. Review the state of any external actions already taken; stopping a run does not undo those actions.
Does a completed task mean its output is approved?
Completion and approval are different decisions. Inspect the deliverable and evidence against your acceptance criteria before relying on the result.
By the Norc team · Product details checked against the Norc project documentation. View the project