Agent Review
Review what an agent changed: watch it work, keep or undo each change, rewind or restore versions, have a fresh agent look for problems, and keep changes reviewable.
Agents can write text that reads well but has a quiet problem: a claim the source doesn’t support, a paragraph that lost its point, a section that drifted from your brief. Hold agent-written work to the same standard as your own. Multilo gives you the tools to see every change, keep only what you want, and ask a fresh agent to look for problems — but the final review is yours.
Watch as the agent works
Agents edit your document in front of you. Watch the first few changes as they arrive: if the agent is heading the wrong way, press Stop or Esc and redirect it early, instead of reviewing a whole chapter of the wrong thing later. For big changes, ask for a plan first, so you agree on the approach before any text changes.
Review every change
A bar above the chat box shows how many files the conversation changed, with Keep all and Undo all. Expand it to see each changed file, with new and deleted files marked.
Keep or undo
| Level | What you can do |
|---|---|
| Everything | Keep all or Undo all for every file the conversation changed. |
| One file | Keep or undo all its changes, open the file, or show its changes. |
| One change | Keep or undo a single change inside a file. |
Undo safely
- If you typed in a file after the AI changed it, undoing the AI’s change leaves your typing alone.
- An undo can itself be undone.
- If a change has moved since the AI made it, Multilo tells you it wasn’t undone instead of guessing.
Go further back
The review bar is for deciding what to keep. To go back in time instead:
- Rewind a chat: choose Rewind to before this message to restore the files, the conversation, or both. See manage sessions.
- Restore a version: right-click a file in the Explorer and choose Restore version… to compare an earlier version with the current one and bring it back. Multilo keeps the last 40 versions of each text file for 14 days. See versions & safety.
Have an agent review the work
A second look catches what the first pass missed. Start a new chat for it, so the reviewer isn’t attached to text it just wrote, and ask for problems — not fixes:
Review the changes to section 3 against my brief in @notes/brief.md. List anything that drifted from the brief, any claim without a citation, and any paragraph that lost its point. Report problems, not style preferences.Then use the reviewers built for the job:
- Claim Check reads every citation, fetches the source, and tells you whether it supports the claim.
- The methodology review agent in the Agents Library checks a Methods section for gaps.
- Your own custom agent or skill can apply the same checklist every time.
Not every finding needs fixing
A reviewer asked to find problems will usually find some, even in good work. Act on the ones that affect your argument or your requirements, and treat the rest as optional.Give the agent checks to meet
Review is faster when the agent already checked its own work against something concrete. Put the checks in the request:
- Sources: “cite only sources from my library, and say so if a claim can’t be supported”;
- Limits: “keep it under 800 words”;
- Requirements: “check the result against the four points in the brief and list anything missing”;
- Evidence: “quote the passage that supports each new claim”.
With strict integrity mode on, the AI also refuses to invent a citation, in every request.
Keep changes reviewable
One section at a time
A change to one section is easy to check; a rewrite of the whole thesis is not. Ask for changes in pieces you can read in one sitting.Review before moving on
Keep or undo each piece before asking for the next, so problems don’t stack up.Commit the milestones
When a chapter is right, commit it to Git, so you always have a version you’ve fully reviewed.
A common failure pattern
The plausible, unchecked draft. The text is fluent, the citations are formatted, and it looks finished — so it doesn’t get read closely. Then a reviewer finds a citation that doesn’t say what the sentence claims. Fluent isn’t the same as correct: run Claim Check, read the key sources yourself, and don’t submit anything you haven’t reviewed.