Watches (Chief keeps an eye on things)

A watch is a standing instruction for Chief to keep an eye on something, so you don’t have to keep checking. Tell Chief in plain language: “keep an eye on the Acme deal”, “tell me if any invoice goes overdue”, or “watch for new leads”. Everything Chief is watching for you shows up on the Watches page (Agents menu). What happens when one fires has changed: a watch is now handled by an agent, and you hear about it only when the agent cannot deal with it.

• Two kinds of watch: a single record (a status change, or any change) or a whole area (anything overdue, or new records arriving). Checks run every 15 minutes by default, or once daily. Each person can keep 20 active watches.
• A standing condition opens one item, not a stream of alerts. When something is overdue, the watch opens a single item on the Watches page. It counts how many times the condition has been seen, carries a state (open, somebody has it, snoozed, resolved), keeps a trail of everything done to it, and closes itself the moment the condition clears. A condition that stays true for a week is one item you can act on, not a week of identical pings.
• An agent is picked from the area you are watching. Overdue tasks and projects go to your Work agent, invoices and bills to Accounting, then CRM, HR, Support, People and Marketing, with Chief taking anything that has no specialist. Nobody sets this up: every watch already has a handler, and the choice is made fresh on every check, so disabling or archiving an agent takes effect on the next one rather than the next time somebody edits the watch.
• The agent proposes real actions against the specific records involved, named by record, and those go through the approval queue you already have. It works with exactly the tools you granted it, at the autonomy level you set, under the same denylist, the same dry run and the same cost ceiling as anything else it does. Agents are approve-first by default and most workspaces leave them that way, so its work waits for a person in Agent approvals. A watch does not fix things by itself.
• A proposal that cannot run never reaches your queue. Every proposal is dry-run against your live data first, and one the dry run says cannot be executed is rejected automatically. It is recorded as rejected rather than deleted, so the run, its cost and the agent’s reasoning stay on the record: an agent repeatedly proposing work it has no tool for is telling you its skills do not match the job.
• You are interrupted only on escalation, for four reasons: no enabled agent covers that area, the run failed, the agent tried three times an hour apart and it is still open, or the watch is set to be handled by a person. That one notification is required and cannot be switched off, because it fires only after automatic handling has already failed.
• On the Watches page every row says who is handling it and what state its item is in. Open one to turn “Let an agent handle this” off, to pin a specific agent under “Handled by” (Automatic re-picks the area’s agent on every run), and to name who it escalates to. You can still pause a watch, resume it, or remove it. Select several to remove them together, or export the list.