Before, the agent followed one behavior everywhere. Now, on the agent's Slack tab, each channel can stay on its default or be customized. Open a channel's settings and you can set whether the agent responds to all messages or only when @mentioned, whether it does the same for thread follow-ups, and whether it responds to people, bots, or both.

Regular users do not configure anything. It is especially useful if you run both a high-traffic general channel and a dedicated Q&A or support channel from the same agent.

Every channel keeps its existing behavior until you choose to customize it, and each setting is independent, so you can be as precise as you want. In a focused Q&A channel, let the agent respond to everything and stay in the thread. In a busier channel, keep it to @mentions only, or filter out other bots so it does not react to automated posts. Even when thread responses are on, the agent only follows up in threads it is already part of, so it will not jump into unrelated conversations.

Start with a dedicated Q&A channel: let the agent respond to all messages and follow up in threads there, then keep busier channels tighter and adjust from what you see.

Previously, pointing a skill at a particular file meant naming it in the instructions and relying on the agent to search for it each time. Now you attach the document directly from a References section in the skill editor, and the skill grounds in that exact file. References are stored by the document's ID, not its name, so they keep working even if someone renames or moves the file.

It's aimed at Knowledge Agent Owners and Admins who want a skill to lean on a specific source, a brand guide, a policy doc, a pricing sheet, rather than whatever a search happens to surface. The people chatting with your agent don't need to do anything, and they don't need their own access to the referenced document. If you don't manage skills, nothing changes for you.

If a referenced document is ever archived or deleted, the skill doesn't fail, it keeps running and flags the reference so you know to update it. You manage everything from one References list in the skill editor, adding or removing references as your skills change.

Head to the skill editor, find the References section, and attach a document your skill depends on. Start with a skill that already leans on one specific source, that's where you'll feel the difference first.

Agents can create and organize collections and folders, move and archive cards in bulk, and run the full draft-to-publish loop on your behalf. Previously, an agent could find and explain your content but every actual change still meant opening the editor yourself. Now you can describe the change in chat — "reorganize these cards into folders by region," "draft a card summarizing this policy" — and the agent does the work.

It's especially built for admins and knowledge managers who spend time on structural upkeep: reorganizing collections, retiring stale content, or building out new spaces card by card. Everyday readers won't see anything change — this only affects what agents can do for the people who manage content, and only within the permissions those people already have.

An agent can only do what you can do — writes are permission-checked operation by operation. Cards are archived rather than permanently deleted, so anything an agent retires is restorable from the UI, and agents can't delete collections at all; that deliberately stays a human decision. Draft edits made from chat sync live into the editor, so if a teammate has the draft open, they see the agent's changes in real time and nobody's work gets overwritten. Agents can even add or remove collaborators on a draft — though a draft's owner can never be removed.

Open a chat with one of your Knowledge Agents and ask it to draft a new card on a topic you know well — it can build with headings, tables, callouts, and collapsible sections, so the result looks like a card your team made by hand. Once you've reviewed and published that first draft, work up to bigger jobs like reorganizing a folder or archiving outdated cards. This is live for all customers today.

Previously, when someone left your workspace, any draft only they owned was permanently deleted. Now those drafts persist, and workspace admins get full visibility into every draft in the workspace. From the Drafts tab in Card Manager, admins can filter between My Drafts, All Drafts, and Has no Collaborators, then claim any draft with one click to edit or publish it.

If you've ever offboarded an employee and wondered what in-progress knowledge walked out the door with them, this is for you. It's particularly valuable for teams with retention or audit obligations, where draft content may need to be produced on demand. Non-admins' experience doesn't change: you still see only your own drafts (plus a new Date Created column for everyone).

Claiming a draft is deliberate and scoped: joining grants you edit and publish rights on that one card only — it doesn't open up broad access to content across the workspace or change what appears in your search results. And the original author isn't cut out: when you claim their draft, they're kept on as a collaborator, so a returning teammate still has access to their own work. Nothing is reassigned automatically; admins claim drafts by choice, per draft.

Head to Card Manager → Drafts and check the Has no Collaborators filter — you might find work worth rescuing.

Instead of clicking through settings screens, just describe what you want: "I want you to be able to answer questions about our leave policy," and your agent drafts the skill with you. You can edit its instructions in plain language, or

If you own a Knowledge Agent — especially one using skills — Operator Mode is for you. It's the answer to the most common question we hear from teams building agents: "It gave a bad answer... now what?" You'll find it on your agent's Chat and Skills pages. (Regular users won't see it; entry requires the Manage agent settings permission.)

Every change your agent proposes comes with an explicit accept/reject confirmation before anything is written, so nothing changes without your sign-off. Once you approve, changes are live immediately — fix a skill after a bad answer, regenerate the response to confirm it's right, and the next person who asks gets the better answer automatically. Skills update in real time on the Skills page, and your Operator Mode sessions are visible only to you.

Operator Mode is live now for all customers. Open one of your Knowledge Agents, enter Operator Mode from the Chat or Skills page, and start with something small — ask it to explain a recent answer, or describe one new skill you've been meaning to build.

Your Knowledge Agents can now connect to Microsoft Work IQ, so they can work directly with these Microsoft 365 apps in real time.

Ask an agent to check your Outlook calendar, draft an email, search SharePoint, or post to a Teams channel, all from Guru chat and all under your existing Microsoft permissions.

An admin enables the Microsoft connections, you sign in once with Microsoft. A Microsoft 365 Copilot license is required to use these connections.

You can now control who can use specific skills on a Knowledge Agent. Restrict any skill to a defined set of users or groups, while keeping the rest of the agent open to its broader audience.

Restricted skills show a lock indicator and are invisible to users who aren't permitted to run them.

This is perfect for layered agents: one general-purpose Knowledge Agent, with sensitive or specialized skills scoped to the right people.

Knowledge Agent owners now have two AI guardrails to lock down their agents.

screens user inputs for attempts to manipulate your agent’s behavior before they reach it.

The new lets you write your own plain-language rules — for example, keeping an agent focused on a specific topic — choose how strictly they’re enforced, and set the message users see when an input is blocked.

Both work across every channel your agent supports (web chat, search, API, and Slack), are off by default, and stay fully under the agent owner’s control.

Knowledge Agent Automations let you schedule a Knowledge Agent to do work for you on a recurring basis. Write a prompt telling the agent what to do, pick a schedule, and the automation runs on its own. Manage everything from the knowledge Agent itself or in the new Automations tab in AI Agent Center. If you can chat it, you can automate it.

I've been using Automations to scrape Slack conversations every night to capture knowledge gaps and FAQ's, saving hours of manual effort!

You can now merge and split cells in any table inside a Card or Page. Right-click a cell to access Merge Right, Merge Down, and Split, and build the more structured tables you've been wanting, with header rows that span columns, grouped sections, and cleaner comparison layouts. Tables pasted in from Google Docs or other editors now keep their merged cells intact instead of flattening on the way in.