How to Review AI Agent Output as a Team

Display is our product. The process below works with other tools too. We name the Display steps where they apply.

The short answer

Agent output fails review for practical reasons, not for lack of effort. The reviewer gets a screenshot, a PDF, or a pasted block in Slack. Comments arrive in three channels. Nobody knows which version the comment is about. The agent never sees the feedback, so a person retypes it into a chat.

A good review fixes four things:

1. One copy. Everybody reads the same rendered page at the same link.
2. One owner. One person decides when the review ends.
3. Comments in place. Each comment sits next to the text it is about, with the version it started on.
4. A loop back to the agent. The agent reads the comments directly and publishes the next version.

Why chat transcripts, screenshots, and PDFs fail as review copies

A chat transcript mixes the result with the prompts, the false starts, and the tool calls. The reviewer must find the final answer first.

A screenshot loses the text. The reviewer cannot select a sentence, click a link, or copy a number into a spreadsheet to check it.

A PDF freezes one version. After the agent revises the document, the PDF in the email thread is wrong, and nothing tells the reader.

A pasted block in Slack or a Google Doc loses the layout of an HTML report or a prototype. Tables, charts, and interactive parts break.

All four have the same second problem. The comments go somewhere the agent cannot read.

The review checklist

Give every reviewer the same list. Copy it into the review request.

Review checklist for agent output
 
1. Sources: Does each claim that needs a source have one? Open two of the links.
2. Numbers: Check three numbers against the original data.
3. Claims: Mark any sentence that states more than the evidence shows.
4. Scope: Does the document answer the question we asked, and nothing more?
5. Decisions: Is each recommendation clear? Who must act, and by when?
6. Tone: Would you send this to the intended reader as it is?
7. Access: Is the audience correct? Check who can open the link.

Three items need a person most. Agents can invent sources and numbers that look real. Agents also write with confidence where the evidence is weak. Ask reviewers to spend most of their time on items 1 to 3.

The process, step by step

Step 1: Publish the output to one link. Ask the agent to publish the document instead of pasting it. In Display, the agent publishes HTML or Markdown and returns a URL. The first publish is version 1. Choose the audience at the same time: Company for colleagues, or Private with named email addresses.

Step 2: Name the owner and the reviewers. Write one request with the link, the owner, two or three reviewers, the checklist, and a deadline. More than three reviewers usually means comments that conflict and nobody to settle them.

Step 3: Comment on the passage, not in a separate thread. Reviewers open the link, select text, and comment. In Display, select text, select Comment, write the note, and select Post. A selection can be up to 1,024 characters. For a change to a whole section, select its heading and describe the change. Each comment records the version it started on.

Step 4: Ask the agent to read the comments and revise. The agent lists the open comments and makes the changes that are clear. It replies to each comment with what changed. It resolves only the threads it addressed. It leaves open anything that needs a decision. Then it publishes the next version at the same link. A prompt for this step is below.

Step 5: Compare the new version with the old one. Open the current link and the earlier version next to each other. In Display, each version has its own address with a /_v/<n> suffix, for example /_v/2. Paid plans keep earlier versions: 10 on Solo and 50 on Pro. If a comment's passage is not in the current version, select Open that version to see the original context.

Step 6: Record the agreement. The owner replies on the page with the version number, for example "Approved v3". Then the owner resolves the remaining threads. Anyone who opens the link later sees what the owner approved and on which version.

The agent prompt that reads the comments

Connect the agent to Display once through MCP. Then paste this prompt into Cursor, Codex, Claude Code, Claude, or ChatGPT:

Find the Display artifact named "Q4 pricing research".
List its open discussions.
For each discussion:
- If the requested change is clear, make it, reply with what changed, and resolve the thread.
- If the change needs a decision, reply with the options and leave the thread open.
Publish the changes as a new version at the same link.
Return the link, the new version number, and the list of threads that are still open.

The Display MCP server includes comment tools. An agent can list, add, edit, and delete comments. It can also resolve or reopen threads, and watch an artifact. The update checks the current version first, so the agent cannot overwrite a newer version by mistake. The setup for each agent is in the MCP server docs, and the comment workflow is in Review comments.

The agent does not read comments by itself. A person asks it, or a scheduled task asks it. Display can send comment notifications to people who watch an artifact.

Access: who reviews from where

Outside guests cannot comment yet. They open the artifact with a one-time password sent to the invited email address. Collect their feedback in a reply or a call, then add it as a comment yourself.

An Owner can restrict Public artifacts or outside sharing for the whole organization. Check that setting before a review that includes a client.

On the Pro plan, colleagues sign in with Google or Microsoft. On every plan, a one-time password by email also works. For the full access model, read how to control access to agent output.

Prototypes and reports need different reviews

A report or a spec gets the full checklist. The risk is a wrong fact or a weak claim.

A prototype gets a shorter list. Check the main path, the empty state, and one error state. Comment on the visible label or heading next to the problem. To review a page of a live website instead, see how to give website feedback to an AI coding agent.

For more on each format, read how to share an AI prototype with your team and how to share an AI-generated report.

Limits of this process

• No diff view. Display shows each version at its own address. You compare two versions by opening both.
• Text comments. Comments attach to selected text. To comment on an image or a chart, select the caption or the nearest heading.
• Guests view only. Invited people outside your organization cannot comment yet.
• The agent waits for a request. It reads comments when someone asks it to.
• Free keeps one version. The Free plan keeps only the latest version. Paid plans keep earlier versions.