谷歌论文：用显式执行状态替代无限增长的Agent历史记录
New Google paper makes a strong case for replacing ever-growing agent histories…
New Google paper makes a strong case for replacing ever-growing agent histories with a small, explicit execution state.
谷歌新论文有力论证了用小型、明确的执行状态取代不断增长的代理历史记录。
Long-horizon agents may need far less conversation history than we give them: this paper finds that replacing the growing transcript with explicit current state cuts token use and often improves accuracy.
长期任务代理可能需要的对话历史远少于我们提供的：这篇论文发现，用明确的当前状态替换不断增长的记录，可减少令牌使用量，并常能提高准确性。
SKILL.state gives the model only the skill instructions, a structured state, and the latest observation. After each step, the reasoning trace is discarded and only a validated state update survives.
SKILL.state 仅向模型提供技能指令、结构化状态和最新观察结果。每一步后，推理轨迹被丢弃，只有经过验证的状态更新得以保留。
That keeps prompt size roughly constant as the task gets longer.
这使得提示大小在任务变长时大致保持恒定。
On a 100-step warehouse task with Gemini-3-Flash, SKILL.state scored 0.94 while using 65,408 tokens. The LangGraph-style baseline scored 0.91 and used 1,062,387 tokens, a 16.2X difference.
在 Gemini-3-Flash 上执行的 100 步仓库任务中，SKILL.state 得分 0.94，使用 65,408 个令牌。LangGraph 风格基线得分 0.91，使用 1,062,387 个令牌，差异达 16.2 倍。
The design does have a boundary: it works only when everything needed later can be captured in the state schema.
该设计存在边界：仅当后续所需的一切都能被状态模式捕获时才有效。
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