Talk to agents

Start with the right agent

Matte is the main workspace agent. Installed apps can add specialists, such as Wells for Inventory Tracking. Your workspace may customize their names and profiles, and only agents available to that workspace appear.

Open an agent from Chat or choose Chat with [agent] on its profile. Start with the outcome you need and enough detail to identify the records involved. For example: “Find items below reorder point at the main warehouse and explain which suppliers we need to contact.”

Agent chat is included on Free with 100 AI credits each month and on paid plans with larger allowances. It also depends on execution being enabled for your workspace. Live work uses credits and the applicable budget. If access or credits block a request, the conversation explains the issue rather than treating the task as completed.

Keep separate chats for separate work

An agent can have several chats with you. Use New chat with [agent] to start a separate topic and History of chats with [agent] to search and reopen earlier work. Open chats appear as tabs. Rename chat gives one a useful title; Close chat closes its tab while history lets you return later.

History can include scheduled work as well as chats you started yourself. Delete a chat from history only when you intend to remove it. Closing a tab and deleting a conversation serve different purposes.

In an agent direct message, you can edit your latest message to correct or refine the request. That edit asks the agent again from that point and replaces the superseded continuation. It is not a rollback of workspace actions the agent has already completed.

Ask about the work in front of you

When a request is launched with page context, it can include the current object, view, filters, selected records, and calendar window. This helps an instruction such as “Review these items” refer to the relevant work. Be specific about the records and intended changes when scope matters.

Depending on its permissions and the enabled capabilities, an agent can search and read records, create views, update records, prepare counts or orders, and use supported inventory, purchasing, fulfillment, calendar, and chat actions. Dedicated human-only flows remain outside its authority.

The agent acts within both its own access and the requesting person’s access. A conversation does not grant access to hidden records or restricted fields. When a request crosses a permission or policy boundary, change the scope or have an authorized person handle that part.

Include a file

Attach an image, PDF, TXT, CSV, or XLSX file and explain what you want the agent to learn or do with it. Images can provide visual context, while text and spreadsheet files can provide lists or reference data.

Agent reading has tighter limits than chat uploading. A turn reads up to the five newest readable files in its conversation context. PDFs must be no more than 10 MB and 20 pages for the agent to read them. Text and spreadsheet extraction is bounded, so a large file may be represented only partially. Files beyond these limits are identified as not read.

Ask the agent to identify the file and relevant contents before relying on a large import or update. Uploading a file by itself does not import its rows into your workspace.

Review work that needs approval

Stock adjustments and transfers, applying or submitting counts, receiving or confirming purchase orders, supplier email, and other consequential actions can require approval. Larger record edits can also require review. The exact requirement comes from the action, workspace policy, and the scope of the request.

A run can finish successfully, stop, fail, or complete only part of the requested work. Activity can show per-record outcomes and links to supporting records. When available, Currently working opens the page the agent is using, so you can watch the relevant work.

1. Read the proposed change shown under Review before making changes. Check the records, quantities, destinations, recipients, or other details that matter to the action.
2. Choose Approve these changes only when the exact preview is correct. Approval applies to that preview and expires after 15 minutes.
3. Follow the resulting response and Activity entries. Open supporting records or receipts to inspect what succeeded.
4. Use Stop remaining work if the agent should stop. Review any completed steps separately; stopping remaining work does not undo earlier changes.

Work with agents in a group or channel

Use New message to include people and agents in a group, or add an installed agent to a channel when your workspace allows agent participation. Mention the agent by name when you want a response. A direct message with the agent does not need an @mention.

While an agent is still working on your request, a follow-up from you in the same conversation and thread can join that active task. This lets you clarify the request without starting over. In shared conversations, another person’s message does not become your authorization for the agent’s work.

When an action needs approval, the requesting person sees the approval controls; other participants can see who the agent is waiting for. Editing a normal group or channel message does not newly trigger an agent just because the edited text includes a mention.

Carry useful preferences forward

Memory can preserve recurring preferences across conversations. My Memory is private to you within the workspace, Agent Memory is shared for one agent, and Shared Memory is available to all agents in the workspace. The agent can show a notice when it saves a memory.

Review saved memories, their scope, and history in the agent’s Memory settings. Current instructions and live workspace records take precedence over saved preferences. Use the agent’s customization settings for its ongoing tone, instructions, role, and permissions, and its scheduling tools for recurring work.

Agent setup, memory, and scheduled tasks →Understand credits and agent budgets →

Related guides