# Troubleshooting

Start with the state you can observe. Record the workspace, study identifier, time and exact error. Keep credentials out of screenshots and messages. Avoid creating duplicate studies while you diagnose an existing attempt.

## Work is not starting

1. Confirm the selected account and workspace.
2. Open the study and inspect its assignee and scope. Check a bound machine in the agent configuration; the manual study composer does not select a machine.
3. Check Machines for the relevant runtime's online status and health.
4. On the execution computer, check that the supported agent CLI is installed and authenticated.
5. Read the study's latest activity and any explicit waiting reason.

Use [connect a machine](/docs/connect-machine/) for daemon status and logs. An online computer does not establish that every CLI or provider credential is usable.

## An attempt reports an error

Read the actual error before retrying. Authentication failures require the relevant account or provider setup; network failures require connectivity; invalid inputs require correction of the input. Preserve partial hand-ins and files so you can distinguish a failed step from work already produced. See [runs](/docs/runs/).

## A connector is unavailable

Open Settings and Connectors. Read its enabled state and explanation, then inspect any required credential configuration. Some sources work without a key; others need credentials. Do not guess a key is optional or paste secrets into study text. Ask an administrator when the required controls are unavailable.

## Review is waiting

Inspect Review and the current stage status. An agent may have handed in its work while a required scoring or review service remains unavailable. In the demonstration study, Grounding hand-ins existed but scoring did not complete. This did not verify later stages or justify marking the study Done.

## Material appears missing

Check Activity before assuming output exists. Files and Notebook can be empty if that type of record was never produced. For a loading error, retry the failed surface and retain the error if it continues.

Ask your lab administrator for help with the study identifier and a credential-free description of the observed state. Use [evidence and results](/docs/evidence-and-results/) to verify the recovered output.
