prompt-as-implementation tradeoffs
Cursor replaces 12K lines of work-tree code with 200-line markdown skill AI EngineerTL;DW
- Cursor replaced 12,000+ lines of work-tree code with 200-line markdown skills using two primitives: agent skills and sub-agents.
- New /worktree and /bestevent slash commands let users spawn isolated agent instances mid-chat without UI dropdowns or version updates.
- Work-tree feature now works with multi-repo setups (frontend + backend separate repos), which was completely disabled before.
- Markdown-based prompts enable server-side iteration—improvements push to users immediately without requiring Cursor version updates.
- Main trade-off: agents now rely on prompting to stay within work-trees rather than hard OS-level isolation, requiring evals and reinforcement learning.
- Best-event comparison UI improved significantly—parent agent now stitches together code pieces from multiple models, not just selecting one winner.
- Cursor training Composer model with RL tasks for work-tree environments to improve staying on track; smaller models like Haiku deviate more frequently.
- Using BrainTrust evals with Cursor CLI to score whether agents work in correct checkout vs. incorrectly in primary repository.
- Feature less discoverable now (power users must know /worktree command exists), but acceptable tradeoff for advanced-only tool.
- Future work: native work-trees UI in Cursor 3.0 agent window and exploring non-git parallelization primitives to avoid disk/speed overhead.
TL;DW
- Cursor replaced 12,000+ lines of work-tree code with 200-line markdown skills using two primitives: agent skills and sub-agents.
- New /worktree and /bestevent slash commands let users spawn isolated agent instances mid-chat without UI dropdowns or version updates.
- Work-tree feature now works with multi-repo setups (frontend + backend separate repos), which was completely disabled before.
- Markdown-based prompts enable server-side iteration—improvements push to users immediately without requiring Cursor version updates.
- Main trade-off: agents now rely on prompting to stay within work-trees rather than hard OS-level isolation, requiring evals and reinforcement learning.
- Best-event comparison UI improved significantly—parent agent now stitches together code pieces from multiple models, not just selecting one winner.
- Cursor training Composer model with RL tasks for work-tree environments to improve staying on track; smaller models like Haiku deviate more frequently.
- Using BrainTrust evals with Cursor CLI to score whether agents work in correct checkout vs. incorrectly in primary repository.
- Feature less discoverable now (power users must know /worktree command exists), but acceptable tradeoff for advanced-only tool.
- Future work: native work-trees UI in Cursor 3.0 agent window and exploring non-git parallelization primitives to avoid disk/speed overhead.
Cursor rebuilt its parallel-agent work-tree feature — isolated git checkouts running best-of-n comparisons — by composing two existing primitives: agent skills and sub-agents. Maintenance burden drops sharply and multi-repo support is gained, but models occasionally escape isolation on long sessions; evals run via BrainTrust and RL training address reliability.