# Give work an always-on AI teammate

Set the job once, then follow its work, decisions, and next steps.

An agent member is a persistent AI teammate in your workspace. Give it an ongoing responsibility, assign or mention it on records, and follow its work alongside your people. Active agents run in Ralti’s cloud even when your browser is closed or your computer is offline. They can wait for the next useful check without losing the responsibility.

## Create your first teammate

- Open Team, then Add member → Agent. Choose an editable job preset or describe the responsibility in your own words.
- In Job, describe the outcome, what good work looks like, and when a person should take over. Name the accountable person.
- In Access, select the sheets and fields it may change, any permitted connected tools, and the people who can review its proposals.
- In When to work, choose the assignments, mentions, selected changes, or calendar schedule that should wake it. Set spending limits.
- Use Preview to inspect the setup and missing information. Preview does not change records or call connected actions.
- Save paused. Review the AI processing permissions and private-source grants, then activate when the setup is ready.

> **Start with a clear responsibility** For example: “Watch open client requests. Identify missing information, propose a next step, and ask the account owner when the answer is uncertain.” A responsibility has an outcome; a single successful check does not automatically mean the business work is complete.

## Know what it is doing

| Profile tab | What you will find |
| --- | --- |
| Work | Ongoing responsibilities, accountable people, completion conditions, blockers, and the next check. |
| Activity | Evidence, checks, proposed changes, review decisions, action receipts, and measured usage. |
| Settings | Instructions, permissions, waking rules, memory, budgets, ownership, and lifecycle controls. |

Assign an agent through the same Person field used for teammates, or mention it in a record discussion. People and agents can share an assignment. Its AI identity stays visible in comments and history, and it preserves the responsible human. Review pending actions in its profile or Team’s Inbox.

## Approve the action you actually mean

A proposal shows the exact change or connected action for review. Approval checks current permissions and source evidence again. Editing the proposal creates a new revision to review. Reversible record changes can run automatically only when explicitly authorized in the setup; human-only rules still apply. Connected actions require exact review. An approved email reply creates a draft in Email; sending is a separate action.

A teammate can read records, propose updates and comments, draft replies from a granted mailbox, use specifically granted connected tools, remember useful context, and hand work to an authorized person or agent. Handoffs keep their original responsibility and spending limits. They do not transfer private-source permissions.

## Choose the right workspace and sources

> **Sheet selection is not confidential read isolation** An agent member can read canonical records across its workspace. Selected sheets constrain its work and permitted changes. Put confidential canonical records in a separate workspace when they must be isolated.

Private mailboxes require the mailbox owner’s grant, filters, viewers, and AI processing permission. Connected tools need their own account permissions and availability. Private results remain visible only to authorized viewers. Removing source access stops future use and invalidates dependent results and memory. Agents do not have arbitrary browser, shell, script, SQL, or generic API-proxy access.

## Keep context useful and cost understandable

Memory is inspectable and source-linked. Correct an entry or forget it when it is no longer useful; business facts still belong in sheet records. Changed or revoked sources can make remembered information unusable.

Workspace and agent daily allowances reset at 00:00 UTC, even when the work schedule uses another time zone. An otherwise affordable check that reaches a daily allowance saves its progress and resumes after the reset. The originating-work budget covers the whole chain of follow-ups and never resets overnight. An oversized request, unknown pricing, or an unresolved charge stays blocked for review. Usage shows measured model cost; provider-managed connection charges are separate.

## Pause, stop, or change course

- Pause prevents new work. Stop requests cancellation of current work and fences later actions. Revoke removes permissions.
- Changing configuration pauses the agent so the updated setup can be reviewed and activated again.
- Resume presents missed work for deliberate catch-up; it does not silently replay every missed event.
- If an external action has an uncertain outcome, reconcile it against the provider before continuing. Ralti does not automatically resend it or claim that cancellation undid an accepted action.

Agent members are distinct from the older scoped helpers under Workflows → Existing helpers. Those helpers keep their own configuration and history. For a new persistent teammate with ongoing responsibilities, start in Team.

