Cursor 3.2 adds /multitask async subagents, worktrees, and GPT-5.5

• Cursor 3.2 turns background agent work into a first-class workflow: cursor_ai's /multitask announcement says queued requests can now break out into async subagents, and cursor_ai's multi-root workspace post adds a single session that can span multiple folders.
• The new agents window is tied to branch isolation, because cursor_ai's worktrees demo shows background tasks running in separate worktrees that can be pulled into the foreground with one click.
• Cursor paired the UI update with GPT-5.5, which cursor_ai's launch post says is live in Cursor at 72.8% on CursorBench, while gdb's note called it the top model on that benchmark.
• Early hands-on reaction clustered around long-running jobs, with jediahkatz recommending async subagents to monitor background work and ericzakariasson saying GPT-5.5 is a strong default for async tasks.

You can read Cursor's CursorBench post, grab Cursor 3.2, and watch the product demos in the /multitask thread and the multi-root workspace clip. One useful detail from the reaction posts is that Cursor employees kept describing GPT-5.5 in terms of monitoring and completing long-running jobs, not just one-shot code generation, in one post and another.

Cursor's new slash command is aimed at the queue itself. According to cursor_ai's main launch post, /multitask can spin queued work into async subagents instead of waiting for the current run to finish.

Two usage patterns show up immediately in the evidence:

• Start parallel runs instead of stacking prompts behind one active agent, per cursor_ai's launch post.
• Re-route already queued messages into parallel work, per cursor_ai's thread.
• Use an async subagent as a watcher for a long-running background job, according to jediahkatz's recommendation.

Cursor tied the new background flow to a rebuilt agents window. In cursor_ai's thread, the company says isolated tasks can run across different branches, then be moved into the local foreground when they are ready to test.

That is a small but real shift in how the IDE frames agent work. Instead of one agent session blocking the rest, the branch becomes the container for parallel runs, and the foreground becomes a place to inspect whichever branch is worth attention next.

Matt Pocock's /github-triage skill screenshot is from a different toolchain, but it lands on the same operational problem: once coding agents can open lots of issues and background tasks, coordination and isolation become product features.

Cursor also added a cross-repo path for agent sessions. cursor_ai's multi-root workspace post says one agent session can target a reusable workspace made of multiple folders.

That matters because Cursor had already made branch-level isolation visible in the agents window. Multi-root workspaces extend the same idea one layer up:

• one session
• multiple folders
• reusable workspace definition
• cross-repo changes inside the same run

The closest adjacent evidence is Matt Pocock's screenshot, which shows a backlog item for multi-repo support elsewhere in the agent ecosystem. Cursor is shipping that shape of workflow directly in the IDE.

Cursor launched the orchestration update alongside a model rollout. cursor_ai's GPT-5.5 post says GPT-5.5 is now available in Cursor, currently leads CursorBench at 72.8%, and is discounted by 50% for a limited window.

The interesting part is how Cursor staff described the fit. jediahkatz's first post says GPT-5.5 "naturally knows how to multitask and monitor long-running jobs to completion," and his later post calls the model a step change for background-task orchestration.

That lines up with the user reaction in ericzakariasson's post, which says GPT-5.5 is a good default for most async tasks. The benchmark pitch is one thing, but the launch copy keeps steering back to autonomous background runs.

The rollout landed with a small date mismatch. cursor_ai's launch post says the 50% GPT-5.5 discount runs through May 2, while the in-product modal captured by kevinkern's screenshot says discounted usage runs through April 29.

So the concrete availability facts in the evidence are:

• GPT-5.5 is live in Cursor, per cursor_ai's launch post.
• Cursor linked the model to its CursorBench ranking.
• Cursor 3.2 is available from the download page, linked in cursor_ai's thread.
• The discount end date appears as May 2 in the launch post and April 29 in the product modal.