Unverified50% confidenceTradeoffExact time
在多轮Agent对话中，每轮都需携带完整上下文，越往后的轮次Token消耗越高
1
Sources
50%
Confidence
Long-term
Relevance
8/19/2026
First Seen
Sources
Related Claims
UnverifiedLoop Engineering因多轮迭代、完整上下文携带和自我检查机制叠加，Token消耗量通常是普通对话的5至20倍75% similarUnverified系统提示词在每一轮对话中都会被重复发送，具有累积的 Token 消耗效应74% similarUnverified当Agent处于循环状态时，每次调用会携带累积的对话历史，导致Token消耗随迭代次数线性甚至超线性增长74% similarUnverified多Agent编排中每层Agent调用都会带来Token消耗和延迟叠加，需要合理的资源规划73% similarUnverified大多数Agent框架采用完整上下文重放策略，每轮调用需将历史对话全量传入，导致Token消耗随循环轮次呈二次方级增长73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/773159API
curl https://kongchang.com/api/v1/knowledge/claims/773159MCP
get_claim(id=773159)