Verified70% confidenceFactExact time

在长链路任务中，若每一步约5%的错误率，经过20步后整体成功率可能跌至不足36%

4

Sources

70%

Confidence

Long-term

Relevance

7/17/2026

First Seen

Sources

GPT-5.6深度解析：分层模型+多智能体，大模型变身工作流引擎

Related Claims

UnverifiedAgent在多步骤任务中存在错误累积效应，每步95%准确率的十步串联后成功率跌至约60%81% similar Verified如果每步准确率为95%，经过20步循环后整体准确率会降至约36%80% similar Unverified长期运行任务若单次成功率仅85%，每月会有4-5次失败；提升至99%则每月预期失败少于1次，才接近企业可用性标准76% similar VerifiedReAct的误差会链式放大：若每步准确率为p，经过n步后整体成功率约为p的n次方，单步95%准确率经过10步后整体成功率约降至59.9%，20步后仅剩35.8%75% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/544859

API

curl https://kongchang.com/api/v1/knowledge/claims/544859

MCP

get_claim(id=544859)