Create and manage an automation agent
Automation agents help your team respond consistently to repeatable workspace events. Unlike chat agents, automation agents do not speak directly to website visitors. They watch for the record events you choose and prepare internal drafts for a person to review. Typical uses include: \- Drafting a first response when a lead arrives. \- Preparing a reply when a customer sends a message. \- Suggesting the next update when a record changes status. \- Drafting an on-demand response from a selected record. ## Create an automation agent 1\. Open **Agents** in your workspace. 2\. Select **New automation agent**. 3\. Choose a suitable starting recipe or begin with a blank agent. 4\. Give the agent a clear name. 5\. Select the object type it will work with. 6\. Choose when it should run. 7\. Select its model. 8\. Write the instructions. 9\. Save the agent. ## Choose when the agent runs An automation agent can run manually or respond to supported workspace events, including: \- A record being created. \- A customer message being received. \- A form submission arriving. \- A portal item being shared. \- A record changing status. Choose the narrowest trigger that matches the workflow. This prevents unnecessary runs and keeps the resulting drafts relevant. ## Write grounded instructions Tell the agent exactly what it should prepare and which record information matters. You can use available variables from the selected object type, together with standard context such as: \- Record subject or title. \- Customer details. \- Custom field values. \- Recent activity. \- The event that triggered the run. A good instruction explains the desired outcome, tone, important facts, and any boundaries the agent must follow. ## Test before enabling Run the agent against a representative record before enabling it. The test shows: \- The resolved instructions. \- The record context supplied to the model. \- The generated draft. \- The selected model and usage information. A test run does not save the draft, change the record, or contact the customer. Use several different records to check that the instructions still work when information is missing, unusual, or incomplete. ## Enable the agent Enable the agent once its trigger and test drafts are reliable. Enabled automation agents can respond to their configured events. Chat agents remain separate and cannot accidentally enter the automation execution path. ## Review generated drafts Automation agents prepare drafts only. They do not make the final customer decision or send a message automatically. A teammate can: \- Review the draft. \- Edit its wording. \- Add missing context. \- Approve and send it. \- Dismiss it when it is not useful. This keeps human judgment at the centre of the customer relationship. ## Draft delivery protection Hub uses a transactional claim to prevent duplicate delivery when a draft is sent. **How it works:** - Before the provider call is made, Hub freezes the reviewed reply, recipient and channel in a single transaction. This prevents a retry or replay from sending the same draft twice. - Once claimed, the draft is finalised together with a deterministic Activity event. - If a provider call succeeds but persistence fails, the saved result is used on retry without sending again. **Unconfirmed outcomes:** If the outcome of a send attempt is uncertain (for example, the provider returned a result but Hub could not record it), the draft is locked and shown as **Unconfirmed**. Operators can: - Check the provider or mailbox to confirm whether the message was delivered. - Select **Record as delivered** to close the attempt without resending. - Select **Close without sending** to end the attempt without a further send. Neither option unlocks the draft for further editing or re-sending. **Edit and dismissal rules:** - Editing and dismissal are only available for unclaimed pending drafts. - Terminal states (delivered, failed, closed) cannot be reset. ## Monitor agent runs Open the agent to review its run history. Each run records useful operational information, including: \- Whether the run succeeded or needs attention. \- The source record. \- Whether the run was manual or automatic. \- Duration. \- Model. \- Token usage. \- Bounded error details. Use this history to identify failing triggers, incomplete record data, or instructions that need refinement. ## Chat agents and automation agents Use a **chat agent** when you want a reusable public assistant that answers visitors from approved object knowledge. Use an **automation agent** when you want an internal assistant to prepare a draft after a workspace event. Use **Ask Hub** when a team member needs live help investigating or acting on the page or record currently in front of them.