Cursor 3 breaks worktree Source tab and cloud-agent submit in early field reports

• A Reddit field report says Cursor 3's cloud-agent submit button did nothing after GitHub was linked, which turns the headline multi-agent workflow into a dead click for at least one paying user.
• The same report also says Cursor's Source tab failed inside a worktree and prompted for repo initialization even while git status showed a live repository in the terminal.
• Cursor's launch post, linked from the main HN thread, positioned Cursor 3 around agentic workflows, worktrees, and cloud execution, which is exactly where the early reliability complaints landed.
• According to the HN discussion summary, the broader reaction split between people who want Cursor to remain editor-first and people who want it to lean harder into multi-agent orchestration.

Cursor 3 shipped with a big workflow swing, and the weird bits showed up fast. You can read Cursor's official announcement, skim the main HN thread, and then compare that pitch with a hands-on Reddit report that hits the exact surfaces Cursor is trying to make central: cloud agents, worktrees, and source control.

Cursor framed the release around a more agent-centric coding loop. The HN submission explicitly calls out the tradeoff between editor-centric development and agent-centric orchestration, plus practical questions around worktrees, model switching, and cloud-agent handoff.

That framing matters here because the early complaint is not about some side panel. It lands on the path Cursor is asking people to adopt first.

synworks' post lists five concrete failures from a second attempt at Cursor 3's multi-agent workflow:

• Clicking submit with the Cloud environment selected did nothing.
• There was no obvious way to open the worktree in an editor window.
• The worktree path had to be copied manually from the UI and reopened from a terminal.
• The Source tab acted as if the repo was missing, despite git state being visible in the terminal.
• Branch selection for a worktree or cloud agent was hard to find or absent.

The sharpest complaint is the deterministic one: the same user questioned why worktree and branch creation appeared to be delegated to the agent instead of cheap local code.

The HN thread reads like a fork in product philosophy. One commenter wants the old model where the developer drives and the agent assists, while another argues that many agents in separate sandboxes feel like the future.

A third view from athoscouto's comment is more specific: editor mode still wins when you need to interrupt, redirect, or swap models mid-flight. Cursor 3 is surfacing that preference split, not hiding it.

The delta summary adds one useful new line from later discussion: a commenter likes cloud agents for throwaway tasks, but avoids having them build foundational code they expect to extend later. That is not a bug report, but it is a practical description of how some early users are already narrowing the tool's role.