⌘K
Change language Switch ThemeSign In
Narrow Mode
Discussion on Inefficient Agent Harnesses
Discussion on Inefficient Agent Harnesses
 ### 宝玉@dotey @Michaelzsguo : 很多浪费其实都来自设计不佳的 harness，小龙虾就是一个例子。 x.com/Michaelzsguo/s…
#### Michael Guo
@Michaelzsguo · 5h ago
罗福莉刚刚写了一篇很不错的文章。即使 Anthropic 正在切断 OpenClaw 这类第三方 agent 对 Claude 订阅的接入，罗福莉依然给整个 AI 生态提供了一个相对乐观的视角。
我很同意罗福莉说的很多浪费其实都来自设计不佳的 harness，小龙虾就是一个例子。它们在每次查询时发起多次高上下文 API 调用，拉低 cache hit rate，又白白烧掉大量 tokens。Anthropic 的这个决定，反而会倒逼整个生态在 context management、prompt caching 和整体效率上做得更好。这样的约束，往往最终会催生出真正的工程纪律，也会逼出新的创造力。
现在用 Claude，或者想办法给 Claude Code “薅到一个划算的 deal”，有时候真的像在菜市场里讨价还价，因为 token economics 已经成了最核心的问题。大家都在焦虑，自己的 tokens 或订阅额度到底会以多快的速度被烧光。我们也都害怕看到那句提示：“Your session will be refreshed at 3 a.m.” 也许，真正的后备方案只有本地模型，而最近 Gemma 4 这样的模型也说明了，这条路正在一天比一天更有吸引力。Show More
1
0
7
4,583
Apr 5, 2026, 6:10 PM View on X
1 Replies
0 Retweets
8 Likes
3,959 Views  宝玉 @dotey
One Sentence Summary
Further discussion on how poorly designed Agent Harnesses impact token consumption.
Summary
Continuing the discussion on Agent efficiency, this post agrees that poorly designed Harnesses (like OpenClaw) are a major cause of token waste. It emphasizes that Anthropic's strategic adjustment will force the ecosystem to improve engineering discipline and context management, thereby fostering new creativity.
AI Score
80
Influence Score 2
Published At Today
Language
Chinese
Tags
AI Agent
Token Consumption
Engineering Practices HomeArticlesPodcastsVideosTweets