🫨简直无敌！DeepSeek Harness被清华微软团队直接“瘦身”：Token消耗砍掉87.93%，成功率反而暴涨14%！
This is a vibe coding post classified by AGK Intelligence as Agent infra & costs (news), kept by the Vibe Coding Radar because it carries real work, not commentary.
🫨简直无敌！DeepSeek Harness被清华微软团队直接“瘦身”：Token消耗砍掉87.93%，成功率反而暴涨14%！ Agent每次工具反馈后都要重新问一遍LLM，哪怕下一步早定死了（写完代码→编译，编译成功→测试）。清华+微软团队用HarnessMM直接把这些“已决定”的步骤缓存下来，跳过重复推理。 核心成果（98个HumanEval任务）： 🔹DeepSeek Harness：Token从96.28k降到11.62k，砍掉87.93%！成功率从71.43%直接飙到85.71% 🔹Codex：成功率稳在90.82%，Token再降29.54%（66.72k→47.02k） 🔹原理：模型一次规划未来命令+精确参数+守卫条件（如“编译成功”），匹配就直接执行，跳过LLM回调。模型权重完全不动 🔹额外：MBPP上也大幅降本，递归自我改进（RSI）让Harness配置持续变强 Agent循环里大量Token浪费在“重新决定已经决定的事”上。这波直接把成本打下来，性能还往上走。 论文：https://arxiv.org/abs/2610.32813 #DeepSeek #AIAgent #Token效率 #清华 #微软 #HarnessMM #大模型降本 #Agent优化 #梁文峰 #梁圣
Posted by SuSu_酥酥👅 (26k followers) 4 days ago · 1k likes · 68.8k views · view the original post on X. Kept by the Vibe Coding Radar as Agent infra & costs. Tools mentioned: arXiv.org.
More vibe coding work like this
- Build an agent on @onkeiki to answer your customers while you sleep. — @orchid_hq
- how do you stop 20+ agents from colliding on the same client work? — @webjuice_ie
- Brilliant paper on optimizing agent environments. — @dair_ai
- Docker is one box. The coding agent can run in others, and switching is one line 🧰 — @marcelotryle
- Holy f*cking sh*t... everyone upgrades Claude and nobody builds the harness — @gippp69
- 1/9 — @TollyLabs
- As AI coding agents get more capable, knowing what they're doing becomes just as… — @EntelligenceAI
- btw in Portals you can now manage your agent CLI versions, update them, and view their… — @imnotchalk
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 7.9k posts from 5k X accounts over the last 14 days, 2.6k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-06 15:18 UTC. Full method.