When Your AI Agent Dies at 3 A.M. - Reliability Engineering for Agents

Keynote
Shanghai
4:30 p.m. - 5:10 p.m.
Venue B(Integrated Building, Room 434 Meeting Room)
Talk Poster
  • Jintao Zhang CNCF Ambassador, Microsoft MVP

    Jintao Zhang — CNCF Ambassador and Microsoft MVP, currently at Kong. A long-time contributor to the Kubernetes and cloud-native communities, and a member of the PyCon China organizing committee. He is also the author of AnkaLoop, an open-source, long-running personal AI agent with tool-execution capabilities. Every incident and line of recovery code in this talk comes from months of this project's production records on real infrastructure.

    zhangjintao

Abstract

Drawing on real incident records from the open-source AI Agent AnkaLoop, five production failure patterns and three fixes borrowed from databases.

Details

At 3 a.m., your AI Agent doesn't crash and doesn't throw an error — it just quietly stops responding. As LLM applications move from demos to long-running services, failure modes shift from "the code has a bug" to "the system is slowly rotting somewhere you can't see": long-lived connections get silently killed, memory gets eaten by leftover files, an Agent corrupts its own config. Drawing on real incident records from the open-source AI Agent AnkaLoop, this talk walks through five production failure patterns and three fixes borrowed from databases — WAL's intent-before-effect, fail-closed recovery, and the checkpoint-vs-journal trade-off — to help you better understand Agents and think about their reliability.