OceanBase details production Agent observability and evaluation practices
OceanBase AI engine R&D lead Zhang Yi detailed AI Vision practices for observing and evaluating production Agents, including layered OpenTelemetry-based trace collection, unified semantics, and end-to-end evaluation centered on traces. He said the workflow filters by scalar fields and task relevance before applying smaller or larger models, while OceanBase stores raw traces and checkpoints and supports analytical and hybrid queries. OceanBase also presented scale and performance figures, but these vendor-associated claims were not independently verified in the report.
The presentation offers a concrete architecture for production Agent observability, spanning telemetry collection, bad-case discovery, real-time evaluation, cost control, and database workloads at very large trace volumes.
今天这篇文章,是 OceanBase AI 引擎研发负责人远智在 “Agent 观测与评测” 技术专场中分享的内容。
会和大家一起从 Codex、蚂蚁阿福等常见的 AI 使用场景出发,聊到规模化的 Agent 实时评估,为大家带来一次面向生产流量的 Agent 可观测实践。
文章会把远智的线下分享,原汁原味地进行还原,干货十足,欢迎大家阅读和留言评论~

大家好,我是张易,花名远智,在 OceanBase 主要负责 AI 引擎相关研发。
刚才在座的老师们聊到了不少 Agent 落地的问题,包括可观测、评测,以及如何进入生产系统、会对生产带来什么影响。接下来,我结合我们做过的工作,和大家交流一下。
#OceanBase#AI Vision#Ant Group#Ant Afu#Codex#Claude#AntCollector#OpenTelemetry