What Agents Can Do

A new connection reads, searches and suggests. An Editor can allow one to write, never over a verified page.

Applies To

liveknowledge

An agent connected to a Knowledgebase can do three kinds of thing, each called a scope: read, suggest and write. A new connection holds the first two. Writing is something an Editor allows, for one connection, in one Knowledgebase.

The Three Scopes

ScopeChip on its cardWhat it lets the agent do
readReadsRead the context pages, the page tree and pages; search; report a question nobody could answer
suggestSuggests EditsEverything read does, plus create a page and suggest an edit to a page
writeWritesEverything suggest does, plus replace a page's writing and move pages
  • A full API token, and a connection your AI tool made through the MCP server, hold read and suggest.
  • A read-only token holds read only, and can never hold more.
  • write is never given by a token. An Editor grants it in the Knowledgebase, and it applies there alone.

The Person Behind the Agent Still Counts

An agent acts for the person or token that connected it, and never does more than that person could:

  • A Viewer's agent only reads, whatever its scopes.
  • A Participant's agent can suggest edits, but not create pages or write.
  • An Editor's agent can create pages and suggest edits, and can be allowed to write.

Allow an Agent to Write

Open the AI Knowledge tab in the footer and find the agent under Connected Agents.

Press the More button (⋯) at the right of its card and choose Allow Writing.

The dialog says what writing allows. Confirm with Allow Writing. The card shows a Writes chip at once.

An Editor's menu on a connected agent's card: Allow Writing.

To take it back, choose Stop Writing from the same menu and confirm. The agent goes back to reading and suggesting. Nothing it has written is undone.

A card with no menu cannot be allowed to write, and says why: its token is read-only, or the person it acts for is not an Editor here.

Never on a Verified Page

An agent that may write still never rewrites a page that is Verified or Review Due. Those are pages a person vouched for, and the one that is due is the page most in need of a person's eye. On either, the write is refused and the agent is told to use suggest_edit instead, so the change waits for a person to accept. It can write directly only on a page that was never verified.

What an Agent Never Does

  • It never verifies a page. A page an agent creates is Never Verified.
  • No MCP tool or CLI command changes the Review Period or Knowledgebase Style. An Editor sets both in the app.
  • It never sees the team's AI key, if it added one.
  • It never reaches a Knowledgebase kept Offline.

Good to Know

  • Allowing writing applies to one connection in one Knowledgebase. The same agent connected to another Knowledgebase still only reads and suggests there.
  • Each confirmation reads "This applies to this knowledgebase only."

Who Can Do What

ActionEditorParticipantViewer
See what each agent may doYesYesYes
Allow an agent to write, or stop itYesNoNo

Was this article helpful?

Still need help?

Try Troubleshooting, or tell us what you were trying to do.

Contact us