Unverified50% confidenceFactExact time

Agent workflows of 5–10 steps can perform extremely stably, but various failure modes emerge in longer trajectories.

1

Sources

50%

Confidence

Long-term

Relevance

9/11/2026

First Seen

Sources

Do Longer AI Agent Trajectories Mean Less Reliability? Four Failure Modes of Long-Horizon Degradation and How to Address Them

Related Entities

AI Agentconcept

Related Claims

UnverifiedAI Agent多步推理中，每一步5%~10%的性能抖动会逐级放大，导致整体成功率断崖式下滑，P95延迟可翻5到10倍71% similar Unverified100行代码的Agent通常缺乏健壮的错误处理、长期记忆、多Agent协作、可观测性和安全沙箱67% similar UnverifiedAgent 评测中由于错误沿任务链累积传播，最终完成率的分差往往比传统 NLP 基准更难弥合67% similar Unverified在多步骤 Agent 工作流中，未被拦截的脏数据会触发下游节点连锁失败67% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/898873

API

curl https://kongchang.com/api/v1/knowledge/claims/898873

MCP

get_claim(id=898873)