--- title: "How to Get an AI Agent to Read Review Comments and Update the Document" slug: /blog/agent-reads-review-comments publishedAt: "2026-10-16" updatedAt: "2026-10-16" author: ott description: "Keep review comments on the published page and give the agent a tool that reads them. The five-step loop with Display MCP tools, what the agent must leave open, and the limit for outside guests. Checked 16 October 2026." --- # How do you get an AI agent to read review comments and update the document? > **TL;DR:** Keep the comments on the published page, attached to the words they refer to, and give the agent a tool that reads them. In Display (display.dev), the docs workspace for agents and people, colleagues select text in the artifact and leave a comment. The agent reads the open discussions through MCP and edits the exact passage. It publishes the next version at the same link, replies with what changed and resolves the thread. Anything that needs a human decision stays open. Each discussion records the version where it started, so the context stays available. People you invite from outside your organization can view artifacts today, but they cannot comment yet. Display is our product. An artifact is one HTML or Markdown file that Display publishes at its own link. We checked the vendor facts on each vendor's own pages on 16 October 2026. --- ## The loop in five steps Your agent does these steps through the Display MCP server. The tool name is in each step. Details: [Review comments](/docs/comments) and [MCP tools](/docs/mcp-server). **1. List the open discussions.** The agent calls `list_comments` with `subject_type: "artifact"` and `short_id`. The default status is `open`. The agent can also ask for resolved threads, or only for threads with new activity since a date. **2. Read the current version and find the passage.** The agent calls `get_metadata`, `search` and `read`. `get_metadata` returns the current version number and the open threads. The agent needs the version number for step 3. **3. Change the exact passage.** The agent calls `edit` with `short_id`, `base_version`, `old_text` and `new_text`. The old text must occur exactly once. The call publishes a new version at the same link. For a larger change, the agent calls the MCP `publish` tool with the same `short_id` and the `base_version`. You can tell it: "Update the page at the same link." If someone else updated the artifact first, the call stops with a version conflict. The agent reads the newer version, applies the change to it and tries again. **4. Reply to the discussion.** The agent calls `add_comment` with `parent_id` set to the root comment ID. A reply uses the anchor of the root comment, so the agent sends no anchor. **5. Resolve the thread.** The agent calls `resolve_thread` with `comment_id`. It uses the root comment ID from step 1. Any signed-in person with view access can reopen a resolved thread. The agent uses `reopen_thread` for that. ### The prompt This prompt from [Review comments](/docs/comments) does all five steps: ```text Find the Display.dev artifact named Quarterly plan. Review its open discussions, make the requested changes when they are clear, reply with what changed, and resolve only the discussions you addressed. Return the updated link and list anything that still needs my decision. ``` --- ## What the agent must leave open Give the agent a rule for what it must not resolve. Add these lines to the prompt: - A comment that asks a question the agent cannot answer from the document. - Two comments that ask for opposite changes. - A change to a number, a price, a date, a name or a legal statement that the agent cannot check. - A comment about tone or design that has more than one valid answer. For these, the agent replies with what it found and what it needs. It does not resolve the thread. The person who owns the decision resolves it. --- ## Get notified about new comments A signed-in person selects the eye icon (**Watch this artifact**) in the Discussions panel to get comment notifications. The MCP tool is `watch`. Watches are for each user. They work for people and for agents with a user credential. An organization-scoped service key cannot watch, and the only comment tool it can use is `list_comments`. Use a user credential for the full loop. --- ## How other tools do the same loop | Tool | Who can comment | Agent can read comments | Comments tied to a version | | ------------------------- | ------------------------------------------------------------------------------------------------ | --------------------------------------------------- | ----------------------------------------------------- | | Vercel preview comments | Team members. Pro and Enterprise can invite external users. All commenters need a Vercel account | Vercel documents a CLI for comments | Each comment is on a preview deployment | | Stacktree client feedback | Clients, without an account | Yes. Its agent can read the feedback and resolve it | Yes | | **Display** | **Organization members with access. Invited outside guests cannot comment yet** | **Yes. MCP tools read, reply and resolve** | **Yes. Each discussion records its starting version** | Google Docs and Notion also have comments that you can resolve. --- ## The limit for outside guests **Outside guests can view but cannot comment.** You can invite a client by email. The client opens the artifact with a one-time password and reads it. The client cannot leave a comment. Collect client feedback in a call or a message, and ask a colleague to post it as a comment. Guest comments are on the roadmap. On Pro and Enterprise, an Owner can also restrict outside sharing for the whole organization. --- ## When another tool is better - **The reader edits the text in a word processor.** Google Docs has co-editing and suggestions. Display comments do not replace them. - **You review a design, not a document.** Figma comments are on the design file. Anyone with view access to a Figma file can comment. - **The prototype is a running app.** Vercel preview comments are on the running preview. --- ## FAQ **Which MCP tools does the agent need?** `list_comments`, `add_comment`, `resolve_thread`, `get_metadata`, `read`, `search`, and `edit` or `publish`. The authenticated server is at `https://api.display.dev/v1/mcp`. On Claude, add Display from [Claude's connector directory](https://claude.ai/directory/display). For other agents, read [Publish with your agent](/docs/publish-with-agent). **Does an update erase the comments?** No. Each discussion records the passage and the version where it started. If the passage is not in the current version, the reader selects **Open that version** to see the original context. **Can the agent post a new comment on a passage?** Yes. A new root comment needs an anchor: the quoted text with the text around it, and a CSS path. Most agent loops need only replies. Read the `add_comment` input in [MCP tools](/docs/mcp-server). **Can a person fix a small error without the agent?** Yes. Anyone who can edit the artifact can fix a line in the page in the browser. Read [Edit an artifact in your browser](/docs/edit-artifacts). **Does the agent resolve a thread by itself?** Only if you tell it to. The prompt above tells it to resolve only the discussions that it addressed. **How long can a comment be?** Through MCP, a comment body can hold up to 10,000 characters. In the browser, a text selection for a new comment can hold up to 1,024 characters. **Related:** [One URL with a version history](/blog/same-url-version-history) · [Review an HTML prototype](/blog/review-html-prototype) · [Claude Code routine for a weekly report](/blog/claude-code-routine-weekly-report) · [AI reports](/ai-reports)