Dagny becomes a shared team assistant — and the engine underneath her learns to keep working long after the page stops waiting.
Team workspaces now have shared chat threads. Everyone on your team sees the same conversation list and the same transcript, anyone can type, and you can watch teammates composing in real time with author names and avatars on every message. Dagny reads the room: address her and she answers, let two people talk to each other and she stays out of it — and when humans are talking amongst themselves, no credits are charged. Concurrent sends are handled cleanly, so two teammates hitting Enter at the same moment never scramble a thread. Shared threads are enabled per team account, and personal chats keep working exactly as before.
Living in a shared room taught us where the chat surface felt cold, so we warmed it up. In team rooms Dagny no longer answers ambient messages with silence — a quick "ok" now gets a short human reply rather than the unsettling impression of being ignored, with genuine person-to-person exchanges still left alone. The typing indicator also appears the instant you press Enter instead of after a pause, so you always know your message landed.
Behind all of it, we moved Dagny onto a new assistant engine built to survive long work. Chat replies stream the same way they always have, but the turn itself is now durable: if your connection drops, your tab times out, or a request genuinely takes minutes of tool work, the answer keeps being produced and is delivered as soon as it's ready — to solo and shared chats alike, with a background sweeper that catches anything still undelivered even if you never write again. Instead of cutting a long turn short with a half-answer, Dagny now tells you honestly that she's still working and follows up with the complete reply. When she needs a decision from you mid-task, she asks with tappable option cards. You can also see exactly what she's doing while she does it: live tool-activity chips during long turns, still there after a reload. Simple messages like greetings get routed to the fast, cheap path again, so a "hey" comes back instantly rather than spinning up the full toolset.
Speed got real attention this cycle. Tools can now run in the same process as the assistant instead of making a network hop on every single call, cutting seconds off each step; the model is nudged to batch independent lookups into one step; conversation history is summarized once it grows long, which shrinks every prompt that follows; and interactive chat now spends less time deliberating before the first word appears. Together these changes attack the two things you actually feel — the wait before Dagny starts talking, and the pauses between her steps. A one-time slower turn may appear the first time a very long-running conversation gets summarized; after that, it stays lean.
The same engine now runs Dagny's background work too — nightly memory curation, hourly check-ins, and scheduled automations — including in personal spaces, which previously accounted for most of the timeouts we saw on long-running jobs. Automations that used to fail against a hard time limit simply no longer have one to hit. Guardrails came along for the ride: hard per-session limits on how much work a single run can do, one-notification-per-run protection so a check-in can never double-ping your Slack channel, and autonomous runs that state a reasonable assumption instead of stalling on a question nobody is there to answer. A handful of smaller repairs round things out: personal space lost a duplicate navigation row and got Search (⌘K) and the Inbox back in its top bar, in-app feedback from autonomous runs now saves reliably instead of failing silently, cost and tool-usage reporting for the new engine reached parity with the old one across our admin views, and internal plumbing was reorganized to make the in-process tool path possible.
Most of the new engine rolls out gradually, workspace by workspace, with an automatic return to the previous behavior if anything goes wrong — you should notice faster, more reliable answers, and nothing else.
Shared threads are rolling out to team accounts now — bring a teammate into a chat and see how it feels with Dagny in the room. As always, tell us what's working and what isn't; we read every note.