Cursor updates Cursor 3 with split agents and 87% fewer dropped frames

• Cursor's 3.1 release thread turned the new Agents Window into a tiled workspace, so you can split panes and run several agents side by side instead of hopping across tabs.
• According to Cursor's voice input demo and the Cursor 3.1 changelog, dictation now records the full clip, uses batch speech-to-text, and adds a hold-to-talk Ctrl+M shortcut with recording controls.
• Cursor's branch-selection demo and its diff-navigation demo both push the product toward mixed cloud and local work: you can choose the branch before launching a cloud agent, then jump from a diff back to the exact file line and edit normally.
• The biggest raw performance number in the release came from Cursor's performance note, which said large streamed edits now drop about 87 percent fewer frames.
• Eric Zakariasson also surfaced an internal workflow that did not make the main announcement: a cloud mode that can babysit pull requests by fixing CI issues, addressing review comments, and resolving merge conflicts.

The Cursor 3.1 changelog is worth reading alongside the earlier Cursor 3 launch post, because the launch sold a unified workspace for local and cloud agents, while 3.1 fills in the missing daily-driver details. You can trace the product arc from the launch's Agents Window overview to the new tiled multi-agent layout, then compare that against the split reaction in the main HN thread, where some developers liked the sandboxed multi-agent direction and others thought the UI was drifting too far from an IDE.

The headline feature is simple and useful: Cursor now lets you split the Agents Window into panes and keep multiple agent sessions visible at once. The official changelog says the layout supports drag-and-drop placement, keyboard navigation, pane expansion, and session persistence across restarts.

That sounds like a small UI tweak, but it is really Cursor cashing out the bigger claim from the Cursor 3 launch post: the product is becoming a workspace for orchestrating several agents across repos, not just a single chat panel bolted onto an editor.

Cursor shipped two fixes for the part of coding-agent UX that usually feels mushy: getting intent in, then getting back to code. Voice input now keeps the full recording before transcription, which the changelog says improves accuracy, and the recording UI adds a waveform, timer, plus cancel and confirm controls.

The diff view also now links straight back to the exact file line. Cursor's demo pairs that jump with normal editor affordances, including manual edits, Tab completions, and go-to-definition, which keeps review from turning into a dead-end preview surface.

Branch selection now happens before you launch a cloud agent from the empty state. Cursor says that replaces the old default of running against your current branch, which should cut down on accidental runs against the wrong target.

Workspace search also gained include and exclude filters. It is a modest feature on paper, but it fits the same pattern as the branch picker: fewer hidden defaults, more explicit control over where the agent is about to act.

Cursor attached an unusually concrete number to the 3.1 cleanup pass, saying large streamed edits now see about 87 percent fewer dropped frames. The same changelog also lists a stack of smaller fixes around long-chat send latency, diff fetches, file tree flicker, conversation stutter, and freezes in threads full of diffs or code blocks.

One underplayed change is that plan tabs now behave more like regular documents. According to the changelog, they now support reliable loading, dirty tracking, reload on plan changes, and Markdown export, which makes plans feel closer to first-class artifacts instead of transient agent output.

The HN reaction to Cursor 3 helps explain why these details matter. In the launch discussion, some commenters praised the multi-agent sandbox model, while others argued the interface was making code secondary and narrowing support for outside model providers.

A late tweet from Cursor engineer Eric Zakariasson exposed a workflow that is more specific than the main release materials. He said teams are already using cloud agents to babysit pull requests by fixing CI failures, handling review comments, and resolving merge conflicts.

That feature sits neatly on top of the Cursor 3 launch post's promise to go from commit to merged PR inside the new interface, but the tweet makes the operational target much clearer: not just generating code, but staying attached to a PR until the noisy cleanup work is done.