The work
An assistant with a job description, not a prompt.
Four things it is expected to do, and one process everything passes through before it touches the outside world.
- The morning briefEvery weekday, before you start
- Today's meetings with the context you'll need for each. The mail that actually matters, separated from the mail that doesn't. What needs preparing, and every conflict it found while looking.
- Inbox triageContinuously
- Reads the day's mail and sorts it: handled, waiting on someone, or genuinely needs you. Routine replies are drafted and held. Nothing is sent without you.
- SchedulingOn request, or when a thread asks for time
- Reads the real calendars involved, finds a slot that survives everyone's constraints and your working hours, drafts the invite, then stops for your approval before anything lands in anyone's calendar.
- Replies in your voiceOn request
- Drafts the response you would have written — learned from what you've sent before and, more usefully, from the edits you made to its earlier attempts.
Before anything leaves
Four checks, cheapest first, and the last one is hostile.
Verification short-circuits: when a cheap check fails, the expensive ones never run. The final layer is a second model prompted to find the fault rather than bless the work, and it defaults to failing the action when it cannot tell.
- 01Structural
- Is the output shaped the way it must be to be usable at all?
- 02Rules
- Does it break something you set — double-booking, working hours, a recipient not in your contacts, a volume cap?
- 03Memory
- Does it contradict something we already know about how you work? If so it is flagged, not sent.
- 04Adversarial judge
- A second model reads it, prompted to attack rather than approve — and fails the action when it is not sure.
The record
When you ask what it did, there is an answer.
Every run writes down what it proposed, what each check returned, what policy decided, what you said, and when. Not a reconstruction after the fact — a record kept as it happened.
It is also what makes the trust ladder possible. Autonomy widens on evidence, and this is the evidence.
What it connects to
Google Workspace and Slack, done properly.
Gmail, Google Calendar and Slack, built natively rather than through a generic connector — because the assistant needs to understand a calendar, not just read one. Access is granted by you through Google’s own consent screen, and revoked there whenever you like.
Microsoft 365 is not supported yet. If that is what your company runs on, tell us — it moves up the list.