Cursor 為 Ultra+ 用戶推出長時間運行代理

• Architecture: Planner/worker/judge system with coordinated subagents capable of spawning nested subagents, creating trees of coordinated work[7][8]
• Execution Model: Agents propose detailed plans requiring user approval before execution, then maintain alignment across hours or days of autonomous work through multiple agents checking each other's work[3][5]
• Model Integration: Custom-built harness integrates various AI models with flexible configuration, tailoring agent behavior to specific task requirements[2]
• Performance: Subagents now run asynchronously with lower latency, better streaming feedback, and responsive parallel execution; previously all subagents ran synchronously[7]
• Output Scale: Demonstrated capability to generate over a million lines of code in extended runs, with pull requests containing 151k+ lines of code merged with minimal follow-up[3][8]
• Task Complexity: Successfully handles multi-file features, large refactors, challenging bugs, authentication system refactoring, platform porting, and chat platform integration[2][6]

Cursor's long-running agents signal a shift toward autonomous software development systems that reduce human oversight requirements for complex engineering tasks. The architecture's ability to coordinate multiple agents and maintain coherence across extended timeframes suggests multi-agent orchestration is transitioning from research demonstrations into production build systems[8]. This development implies software teams should prepare for agent-driven workflows, with potential implications for developer productivity metrics, code review processes, and the role of human engineers in software development cycles. The emphasis on self-driving codebases indicates Cursor's strategic direction toward systems requiring minimal human intervention for larger project scopes.