Unverified50% confidenceTradeoffExact time
ReAct's reasoning chain continuously consumes context tokens; for an agent executing 20 steps, the accumulated reasoning history may consume tens of thousands of tokens.
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
上下文工程：让AI Agent不再"胡言乱语"的系统方法
Related Claims
Unverified典型的Agent框架如ReAct、AutoGPT、LangGraph采用感知-推理-行动的迭代模式，上下文随轮次累积导致后期每轮Token消耗呈线性甚至指数增长76% 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% similarUnverifiedMost Agent frameworks adopt a full context replay strategy, requiring the entire conversation history to be passed in with each call, causing token consumption to grow quadratically with the number of loop iterations.75% similarUnverified推理-行动循环（ReAct Loop）可以持续多轮，AI分析状态、决定调用工具、获取结果后再规划下一步直至任务完成71% similarUnverifiedA typical ReAct-paradigm Agent may need to go through 5–20 loops to complete a complex task, with each loop involving at least one LLM call.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/607419API
curl https://kongchang.com/api/v1/knowledge/claims/607419MCP
get_claim(id=607419)