Unverified60% confidenceFactExact time
In each ReAct cycle, the complete history context must be re-input, causing the context window to grow linearly with the number of iterations.
2
Sources
60%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Mindwalk：用3D代码地图回放AI编程代理行为轨迹
hackernewshackernews7/12/2026
Related Claims
UnverifiedReAct 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70% similarUnverifiedFor conversational agents, without caching, the Nth-turn request must reprocess all of the history of the previous N turns, and cumulative input grows quadratically with the number of turns; with prefix caching, each turn only needs to process the newly added part, and cumulative input grows linearly with the number of turns.69% similarUnverifiedThe cost of ReAct is that each loop adds one model call, with latency and token consumption growing linearly with the number of steps, so real systems typically set a maximum number of iterations and early-termination conditions.69% 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68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/611091API
curl https://kongchang.com/api/v1/knowledge/claims/611091MCP
get_claim(id=611091)