Catch-up triage and team briefings

Agents propose faster than humans can review. With teammates watching your inbox and CRM on a schedule, the pending-actions list grew quicker than anyone could work it — and the approval UX was the bottleneck: a wall of cards, popups on popups, raw JSON where a decision should be. This release rebuilds both halves of the human loop: clearing the queue, and catching up on what your teams know.

What's new

• A queue that stays clean. Every proposed action now carries a dedup key, auto-derived when the agent doesn't supply one. A re-surfaced action — the same owed reply proposed again on the next check — updates the pending item in place instead of stacking a duplicate. Proposals can also carry an expiry, so stale suggestions drop out on their own instead of rotting in your queue. (Only pending items dedupe; a decided action can always be re-proposed later.)
• Focus-mode review: one flow, no popups. This started as a Slack-style catch-up stepper over the queue; by the end of the window the whole review page became that flow. You see one action at a time as a full editable detail — rationale, confidence, draft fields — decided with Approve, Save-for-later, Skip, or Reject. An up-next rail replaces the enormous list: the list is navigation, the detail is the surface. There's a Back button, and internal mechanics like skill drafts sit in a collapsed "Other approvals" section.
• Human headers, and everything is steerable. Every item leads with what's actually being approved, in plain language — "SEND email → tina@…", "Update HubSpot deal record" — with the raw payload demoted to a drill-down. Every field is editable in place: your version is what runs. And every item (not just emails) gets a steer box with Rewrite with AI — type a hint, get a rewritten draft back for review before anything executes.
• Typed decision signals. Approve, edit-then-approve, reject, skip, save, and rewrite are each recorded as distinct signals, scoped to the individual, the action type, and the workspace — so tone and confidence learning can tell your clean approvals from your rewrites.
• Team-scoped briefings. Every brief now belongs to a team. The briefings page gets tabs — the workspace rollup first, then each team — plus a previous-briefs history explorer and a Regenerate button that runs the owning team lead in the background, with a live "Assembling…" state and the fresh brief auto-landing when ready. No manual reload.
• Chat with a brief, scoped correctly. Ask a question from any brief and you land in chat with that team's lead (the workspace lead for the rollup), brief and question already in the thread.
• freshen_source: fresh data at generation time. A new tool runs a fast incremental sync of a connected source — only what changed since the last watermark — so a brief or a "did they reply?" answer reflects up-to-the-minute data even when the scheduled sync lags. It degrades gracefully, and says so when it can't reach the source.

Why we built it

The review queue is where agent leverage either compounds or dies. If approving is slower than proposing, the queue becomes a guilt list and people stop looking — "human in the loop" quietly becomes "human out of the loop." The fix wasn't a bigger list; it was making each decision cheap: one item, plain language, edit anything, four verbs, next. And because every one of those verbs is now a typed, scoped signal, the act of triaging is the training data — the system learns your tone and its own confidence from how you decide.

Briefings had the mirror problem: one global brief can't serve a workspace running several teams. Scoping briefs to teams — with a rollup synthesized from them — means each lead reports on its own lane, and you read as wide or as narrow as you need.

Where to find it

• Review queue: /dashboard/review — focus mode is the page.
• Briefings: /dashboard/briefings — rollup and team tabs, history, Regenerate, and the chat pill on every brief.
• Docs: Actions covers propose → gate → execute and the queue behavior.
• Agent tools: publish_briefing, get_briefing, refresh_briefing, and freshen_source are available to agents today.