UC Berkeley与微软论文:外部控制器降低编码Agent成本
New UC Berkeley and Microsoft paper cuts coding-agent token costs by 18.4% to 30…
编码Agent长期面临Token成本高昂痛点,这篇论文提供了一套可复用的外部控制框架,实测降本超30%且提升成功率,做Agent工程落地的同学值得参考其架构思路。
New UC Berkeley and Microsoft paper cuts coding-agent token costs by 18.4% to 30.9% using an external controller that learns from past runs.
加州大学伯克利分校与微软的最新论文表明,通过一个从过往运行中学习的外部控制器,编码代理的 token 成本可降低 18.4% 至 30.9%。
by injecting learned facts, pruning unused tools, and stopping failed loops early.
该方法通过注入学到的事实、剪枝未使用的工具,并尽早终止失败的循环来实现。
Coding agents are billed by the token, and each turn resends the growing history, so long projects get expensive. Token efficiency is usually measured on isolated tasks, missing the cost that builds up across a sequence of tasks in the same codebase.
编码代理按 token 计费,且每轮都会重新发送不断增长的对话历史,因此长期项目成本高昂。Token 效率通常仅在孤立任务上衡量,忽略了在同一代码库中一系列任务累积产生的成本。
TokenEconomizer sits outside the agent as a black box, changing only prompts and tool settings. From past runs it learns which facts to inject, which unused tools to drop, and when to stop a failing loop.
TokenEconomizer 作为黑盒位于代理外部,仅更改提示词和工具设置。它从过往运行中学习应注入哪些事实、丢弃哪些未使用的工具,以及在何时停止失败的循环。
Across Claude Code, Codex, and GitHub Copilot CLI on 277 SWE-Bench tasks, it cut mean per-task cost 18.4% to 30.9%. It improved Pass@1 in 4 of 5 agent-model settings, and savings held in every repository.
在涵盖 277 个 SWE-Bench 任务的 Claude Code、Codex 和 GitHub Copilot CLI 上,其将每个任务的平均成本降低了 18.4% 至 30.9%。它在 5 种代理-模型组合中的 4 种提升了 Pass@1 指标,且节省效果在所有仓库中均保持一致。
Measure agent cost across a whole project, not per task, and use an external controller that adapts to each repo, not fixed plugins.
应针对整个项目而非单个任务来衡量代理成本,并使用能适应每个仓库的外部控制器,而非固定插件。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力