UC Berkeley 研究:不改模型、只改 harness 可省最高 2.6 倍 agent 成本
UC Berkeley 这篇论文说清了一件事:降 agent 成本不一定非换模型,改 harness 层就能差出 2.6 倍。他们还开源了 Agency 这层中间件,写一次优化就能套在 Claude Code、Codex 上。跑 agent 产品、被 API 账单压着的团队都该看看,先删重复上下文再说。
UC Berkeley 研究人员分析了 agent 运行轨迹,发现提示词和工具等 harness 层面的改动对准确率影响很小,但在同一模型和任务上可让成本相差最高 2.6 倍。团队总结出五种降本方式,核心思路都是减少每次 LLM 调用读取的上下文长度,或减少完成任务所需的调用次数。
基于这些发现,他们构建了 Agency,一个介于 Claude Code、Codex 等现有 harness 与模型之间的优化层,一次写好的优化可以跨 harness 复用。五个概念验证技术实现了最高 35% 的成本削减,准确率损失微小。
对开发者的直接启示是:agent 账单持续增长时,先检查重复的上下文读取和多余的工具调用,再考虑换模型。
Great practical paper on optimizing harness costs.
For lower agent costs, you want to change the harness and keep the model.
UC Berkeley researchers found that harness changes such as prompts and tools barely move accuracy, but they change cost by up to 2.6x on the same model and task.
They analyzed agent traces and found five ways to cut cost. Each one reduces how much context an LLM call reads or how many calls a task needs.
Then they built Agency, a layer that sits between existing harnesses like Claude Code and Codex and the model. An optimization written once works across harnesses.
Five proof-of-concept techniques cut cost by up to 35% with marginal accuracy loss.
If your agent bill keeps growing, look at repeated context reads and extra tool calls before switching models.
https://t.co/aaLRm7tOLw
Chat with Paper: https://t.co/FQgxm3IHgu
本条情报归属 AITOP「论文」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:Agency、Claude Code、Codex、agent,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明:本文原始线索采编自「DAIR.AI」(@dair_ai),经 AITOP 多维评分管线结构化提炼。