When an agent makes an unexpected change, first find the affected item, the change, and the surrounding context. History is most useful when you have a concrete question to investigate.
Suppose a project memo now has a different deadline. You want to know what changed, whether it came from a person or an agent, which source justified it, and whether later work depends on the new date.
Start with the review surface
In a rich-text document, inspect the suggestion. In a spreadsheet, inspect the changed cells and the Unreviewed changes panel. For a canvas, review the source and baseline comparison. For a File, compare the relevant versions.
These views keep the investigation tied to the artifact. They can be more useful than a broad conversation about what the assistant intended to do.
Use history to establish the sequence
The current Guide describes a revision timeline with actor, time, action type, affected locations, and previews where available. Agent-authored entries can show the agent, the person it acted for, and an optional stated reason.
A reason is the agent's explanation. Treat it as something to verify against the task and sources. A confident explanation does not prove that the edit was correct or authorized for distribution.
History can help answer whether the date changed in an agent pass or a later human edit. It can also show that the apparently simple change belongs to a larger session that altered several sections.
Separate content history from authorization records
Content history describes how an item changed. Authorization records describe access decisions and related activity. A conversation transcript describes the task context. These are different records, and one does not necessarily contain the complete story of another.
Skybits documents agent access records as well as content history. The existence of logging alone does not establish a permanent, complete record of every event. Confirm the coverage, retention, and retrieval of the records required by your workflow.
For a workflow requiring formal evidence, ask which records are available, how they are retained, whether they are protected against alteration, and how they can be retrieved. A visible AI badge is useful operational information; it is not an independent compliance attestation.
Recover with attention to concurrent work
When a narrow rejection or correction can solve the problem, review that option before restoring the whole item. Restoring a prior state can also affect useful work added since that revision.
The current Guide describes full-item restoration and notes that agent restore permission can replace the current item, including comment threads, with a prior revision. Check with collaborators when their recent work may be affected.
Recovery repairs the workspace artifact. It does not retract an export, message, or provider request already made from the changed content.
Record the resolution in useful terms
For an internal follow-up, record the affected item link, approximate time, relevant revision, expected result, observed difference, and chosen correction. Avoid pasting credentials or unrelated document content into a support request.
If the cause was an unclear task, improve the source and stopping point. If it was excessive authority, reduce access. If the behavior conflicts with documented product controls, raise the specific discrepancy with the team.
Next step: Pick one recent agent change and trace it from review to history. Confirm that your team knows how to inspect and recover it before assigning a larger task. See Skybits Help.