The Agent is the fifth tab of the Chat & Tasks panel — an AI assistant that works on your brand's live PLM data, next to the team's messages and tasks. Ask it a question and it answers from what is in the system right now; ask it to change something and it rehearses the change first, then shows you exactly what it is about to do and waits for your go-ahead before anything is written.

Open Chat & Tasks from the speech-bubble icon in the top bar and pick the Agent tab, or jump straight there with ⌥ ⌘ A (Alt A on Windows and Linux). The command palette (Ctrl/⌘ K) has the same action under Open agent chat. The panel opens beside the page you are on, and that page — the product or article in front of you — is attached to your message as an Auto-linked context chip, so "this article" means the one you are looking at. Remove the chip with its × if you don't want it included.

Work is kept in threads, shown as tabs under the Agent heading. + New starts a fresh one; each thread is named after its first message and keeps its own history, so a question about samples and a batch of price edits don't get mixed up. Type in the box at the bottom — Enter sends, Shift + Enter adds a line — and use the paperclip to attach a file for the agent to read.

While the agent is thinking, the composer turns into a queue: whatever you type is held and read when the current turn ends (Queue), and Stop in the thread header ends the turn early.

Ask in plain language — which colourways a product has, which sizes still lack an EAN, what a round's status is. The agent reads from the PLM and the brand's configuration and answers with the numbers and names as they stand. Above each answer sits a collapsed steps line (Read from the PLM · Looked up the API · Read the brand configuration); expand it to see what the agent did to get there, including any step that failed.

Ask for a change — set a comment, add a CC address to a sample round, fill a field — and the agent works it out in the open. It looks up the right operation, rehearses it as a dry run against the real data, reads the current value, and then stops on a highlighted step that says what it wants to do:

Allow lets that change through; Deny stops it and the agent says so. Nothing is written until you decide. Once the change is made, the step turns into Made a change and the answer confirms the new value, so the thread doubles as a log of what was done.

If you trust a thread — a long batch of edits you have already checked the pattern of — switch on Auto-approve writes in the thread header and the agent stops asking for that thread. The switch is per thread and off by default.