Unverified50% confidenceFactExact time
ReAct 每次循环产生 Thought-Action-Observation 三元组，会被逐步拼接进上下文窗口，多轮循环会快速消耗模型的 Token 配额
1
Sources
50%
Confidence
Long-term
Relevance
7/6/2026
First Seen
Sources
Hermes Agent 安装实战：3分钟跑通AI智能体完整流程
bilibili比特光锥_BightCone7/3/2026
Related Claims
UnverifiedReAct循环表现为Thought→Action→Observation的三段式格式，其局限性在于长任务中上下文窗口快速消耗以及可能陷入推理循环79% similarUnverifiedReAct Loop每次迭代包含Thought、Action、Observation三个阶段78% similarUnverifiedReAct's workflow consists of three repeating parts: Thought (reasoning), Action (tool call), and Observation (tool result).75% similarUnverifiedReAct范式通过将思考（Thought）→行动（Action）→观察（Observation）三元组交织在同一上下文中构建闭环反馈控制系统75% similarUnverified在Agent应用中，每一轮推理-行动-观察循环都需要将完整历史上下文重新输入模型，导致Token消耗随任务复杂度呈非线性增长75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/114979API
curl https://kongchang.com/api/v1/knowledge/claims/114979MCP
get_claim(id=114979)