Concurrent Local Coding Agents – My opinionated, flexible version of Cursor 2.0 (xxchan.me)

Cursor 2.0 introduces a concrete push toward local, concurrent coding agents by adding worktree-based local agents, a multi-agent “race” mode (useful for pass@k-style multi-generation selection), and an Agent Chat–first UI alongside its existing Background Agent. This validates a practical hybrid: remote agents give independent parallelism and convenient task assignment, but suffer vendor lock-in, token-cost volatility, and last-mile friction when you need local control. Cursor’s new features acknowledge those tradeoffs, but the release still falls short on key ergonomics — no quick preview/preview-centric UI and limited ability to run dev commands inside a single worktree — so local workflows remain incomplete. The author reports a pragmatic evolution from chaotic single-workspace parallelism to disciplined concurrency with git worktrees (inspired by matklad) and CLI tools like xlaude, Conductor, and a custom AgentDev. Worktrees let multiple local branches run isolated agents and preserve local dev comforts (pnpm dev, VS Code, gh pr view). Pain points persist: reviewers become bottlenecks, context-switching breaks focus, and merging many tiny PRs is tedious. The takeaway for the AI/ML community is twofold: local concurrent agents + worktrees materially lower cost and increase flexibility today (especially using first-party CLI models), but UX and verification tooling (previewing results, running commands per worktree, PR-like context) need polishing before this pattern scales widely.