The Confluence Knowledge Base That Maintains Itself

Every company wiki dies the same way. Somebody writes the onboarding page, it is excellent, and then the process changes and nobody updates it. Six months later new hires read it, follow it, and get it wrong. So people stop trusting the wiki and start asking in chat instead, which is exactly the cost the wiki was supposed to remove. Documentation does not rot because writing is hard. It rots because maintaining it is nobody's job.

Confluence is connected in AgentTeams, and that makes it two things at once: the place your agents look up answers, and a set of pages your agents keep current. The overview lives on the AI agents for Confluence page. This article covers what is actually possible, how the connection works, and the setups worth stealing.

Reading: the wiki becomes answerable

• Search the whole wiki. Your agent can search across spaces by text, title, author, space, or last-modified date, so questions like "what did we decide about refund policy" hit the page instead of a person.
• Walk spaces and pages. List the spaces the connected account can reach, open a single space, and list the pages inside it, including archived and draft ones when you ask for them.
• Read a page in full. The body, not the preview, in the editable format, the rendered view, or the structured form, whichever the task needs.
• Read the discussion. Footer comments on a page are often where the real decision lives. Your agent reads that thread too.
• Open the attachments. It can list what is attached to a page and pull the text out of PDFs, Word files, spreadsheets, and plain text, then hand the file straight to another tool without you downloading anything.
• Know who is who. Look up the connected account, search people by name, and list the groups on the site, so a page can name the right owner instead of "the team".

Writing: the wiki stops going stale

• Create pages. In the right space, under the right parent, published or left as a draft for a human to finish.
• Update pages. Your agent reads the current version first, rewrites what changed, and saves it with a version note, so the page history says what happened and why.
• Comment. Post a footer comment on a page, which is the polite way to flag "this section is out of date" without silently overwriting somebody's work.
• Label and unlabel. Keep the tags that drive your page trees and filters accurate as content moves around.
• Trash a dead page. Send a genuinely dead page to the trash instead of leaving it to confuse the next search.
• A generic confluence_api passthrough for the rest: blog posts, version history, moving a page under a new parent, inline comments, and anything else we have not wrapped yet.

Creating, updating, deleting, and commenting all go through your approval flow until you say otherwise. The realistic starting point is not an agent rewriting your wiki overnight. It is an agent that proposes the edit, shows you the before and after, and waits.

It answers when you pull it in

@mention your agent on a page, or leave a comment on a page it watches, and it responds there within about a minute. It reads the page, the thread, and whatever else it knows about the topic from your other tools, then replies in the comment thread where the question was asked. That is usually how the first real conversation with the wiki happens: someone comments is this still true? and gets an answer instead of silence.

How the connection works

Open your agent's Tools tab, click Connect on Confluence, and sign in to your Atlassian site as the account the agent should act as. Every page, edit, and comment is attributed to that account in Confluence's own history, and the agent sees exactly the spaces and pages that account can see. Give it access to the spaces it should maintain and leave it out of the ones it should not touch. Revoking the account in Atlassian ends everything immediately.

Setups worth copying

• The stale-page review. Once a week the agent finds pages in a space that have not been touched in six months, reads them against what it knows is true today, and comments on the ones that no longer match, with the specific line that is wrong.
• Decisions that get written down. A decision gets made in chat or in a ticket. The agent drafts the page or the update to the existing page, in your house style, and asks the person who made the call to approve it. The decision stops living in someone's memory.
• Support answers that feed back. When your support agent answers the same question for the third time, the documentation agent proposes the page. Pair it with the knowledge base your agents learn from and the loop closes on itself.
• Release notes with the work attached. Pair Confluence with Jira and the same agent that watched the tickets close writes the page that explains what shipped, with the issues linked.
• Onboarding that is actually correct. New hire asks a question, the agent answers from the wiki, and if the wiki was wrong it proposes the fix in the same breath. Every new hire makes the docs better instead of just discovering they are broken.

Common questions

• Will it edit pages without asking? Not unless you decide it should. Page writes and comments sit behind approval by default, and you can keep them there permanently for spaces where nothing should change without a human.
• What about restricted pages? The agent sees what the connected account sees. Confluence does not currently let a connected account like this read page-level restriction settings, so treat access as something you grant deliberately per space rather than something the agent negotiates.
• Will it mangle our formatting? No. Headings, lists, bold, and links are written in Confluence's own format, so pages look like pages, not like pasted text.
• Can it read our attached files? Yes, for common document and spreadsheet formats. Very large files come back shortened, with a note saying so.
• Do we need a separate wiki for the agents themselves? No. Confluence can be the source your agents read from, and they carry that context into every other tool they work in.

Where this fits

Confluence sits with Notion, Tettra, and the rest of the knowledge side of the catalog. The interesting part is not that an agent can write a page. It is that the agent writing the page is the same one that answered the support ticket, watched the ticket close, and sat in the channel where the decision was made. A wiki maintained by the person who was there beats a wiki maintained by whoever has time on Friday.

Ready to try it? Open your agent's Tools tab and click Connect Confluence. New to AgentTeams? Start with the first-agent onboarding guide.