Unverified50% confidenceFactExact time
Agent Reliability将评测器自身的执行失败与Agent的失败严格区分，并将测量健康度与Agent可靠性分开追踪
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAgent Reliability的核心思路是把SRE（站点可靠性工程）的可靠性度量方法迁移到AI Agent评测中78% similarUnverified衡量Agent长程可靠性的核心指标包括任务成功率、工具调用准确率、恢复率、单次成功任务成本和人工干预次数77% similarUnverifiedAgent Reliability的基础包零强制运行时依赖，并内置确定性评测器76% similarUnverifiedAgent可靠性验证方法论包括：每个任务重复运行10次以上、以程序化方式检查最终状态、对每次工具调用按schema校验、统计高负载下被截断的调用数量74% similarUnverified由于Agent输出具有不确定性，其测试需要设计评估指标（如准确率、任务完成率、延迟）并通过批量测试用例验证稳定性73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/916847API
curl https://kongchang.com/api/v1/knowledge/claims/916847MCP
get_claim(id=916847)