Smarter tool use: your assistant now plans before it acts

A batch of improvements shipping today, all aimed at one thing: when you connect your apps, the assistant should actually use them well.

Planning before doing

Any reply that involves your connected apps now starts with a fast planning step. The assistant reads the real list of actions it has, checks what it remembers, and drafts a short plan — then executes. You can watch the plan appear in the Deliberation trace above the reply. The difference is night and day on multi-step requests: fewer wrong turns, fewer "I can't do that" answers when it plainly can.

Execution plan
1. jira_search_issues — open tickets assigned to James
2. confluence_search — pages mentioning those tickets
3. Summarise by priority
 jira_search_issues
 confluence_search
You have 4 open tickets — 2 urgent. ONE-42 blocks the release and has a Confluence spec that changed yesterday. Want the diff?
 2 tools · planned first

The whole catalog, not a slice

Connected apps can expose a lot of actions — Jira alone has over a hundred. Previously the assistant worked from a curated subset, and occasionally the exact action it needed wasn't in it. Now it can search the full catalog mid-conversation and run anything it finds, with the same safety rules: in Safe mode, actions that make changes stay blocked.

Also in this release

• Move conversations from the toolbar — a new button opens a picker of every model that can take over your chat, with incompatible ones explained rather than hidden. Attached agents travel with the conversation.
• Copy your own messages — hover any message you sent.
• Mentions look like mentions — @-mentions render as badges in the composer and in your sent messages.
• Fixed: replies that finished while you were on another page now appear automatically when you come back.
• Fixed: memory syncing interruptions some users saw this week — writes now retry through a fallback path automatically.