A teammate with follow-through.
A clear job. Shared context. You stay in charge.
- 01Context checked
Reviewed the assigned work and linked evidence.
- 02A useful next step
Prepared a record update for Maya to review.
- 03Following through
Waiting for the next change. The responsibility stays open.
Start with a responsibility worth keeping
Create an agent from a job preset or your own instructions. Define the objective, what completion means, the human owner, and when the agent should check. Preview its scope and missing evidence, save it paused, then activate deliberately.
- Editable presets for waiting work, support triage, project coordination, time quality, and delivery review
- Assignments and mentions alongside people, with a visible AI label
- A responsibility stays open until it is actually resolved—not merely because a check finished
Follow-through that survives a closed laptop
Activated agents work on the server. They respond to permitted assignments, mentions, selected changes, and schedules, and can arrange their next useful check. An unchanged source does not need another model call just to stay under watch.
- Durable responsibilities and scheduled follow-ups
- A Work view with owners, next actions, next checks, and blockers
- Deliberate catch-up when work was missed during a pause
Evidence before action
Inspect the records and other permitted sources behind a proposal. Review the exact change or connected action before approving it. Current permissions and source freshness are checked again when the action is applied.
- Exact sheet and field write permissions, named reviewers, and approved connected tools
- Private mailbox access requires the mailbox owner’s grant
- Native email reply proposals create a reviewed draft; sending is a separate action
Stay close without watching every move
Use the profile and Team Inbox to see what needs attention. Inspect or correct memory, revise instructions, transfer supervision, and choose whether to pause new work, stop current work, or revoke access.
- Versioned instructions and source-linked, expiring memory
- Workspace, agent, and originating-work spending limits
- Uncertain external outcomes stay blocked for a person to reconcile
A little more clarity.
Are agents always running a model?
No. Active agents remain available to their responsibilities, but checks are admitted only when there is useful work. Waiting releases compute. Daily limits can defer affordable work to the next reset; unresolved charges or actions require review.
Can I restrict an agent to reading one sheet?
Not as a confidential access boundary today. Agents read canonical records within their workspace membership. Selected sheets narrow their work and permitted changes. Use separate workspaces when records must be isolated; private email and connected sources need separate grants.
Will a quality agent approve work or change recorded time?
Quality presets support review, not automatic acceptance or client commitments. Time-quality review needs detailed session evidence; timer totals alone are insufficient. It must not rewrite hours, billability, attribution, or invoices.