推文

@DracoVibeCoding · 2026-10-12 20:38

Coding agent benchmarks grade the wrong job. JetBrains checked 4,782 real IDE sessions. SWE-Bench Pro: 93.8% features + bug fixes. Real long sessions: well under half. The rest is explain, plan, review. Ask in steps like a human: ~2x cost, no clear gain. #CodingAgents

曝光 167 · 评论 3 · 点赞 0 · 书签 0 · 曝光/时 14.96497532985793