Unverified50% confidenceFactExact time
Agent 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
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAgent 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78% similarUnverifiedAgent Reliability的基础包零强制运行时依赖，并内置确定性评测器70% similarVerifiedAI领域模型能力的可信评估依赖标准化基准测试，常见评测集包括MMLU、HumanEval、MATH、GPQA69% 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69% 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69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/916841API
curl https://kongchang.com/api/v1/knowledge/claims/916841MCP
get_claim(id=916841)