We built a durable agent that debugs durable agents

Almog Baku, CEO and co-founder of Kelet AI, explains how Kelet, built on Temporal, addresses the challenge of diagnosing AI quality failures that do not manifest through traditional stack traces but rather emerge from patterns across multiple sessions. Kelet is designed to continuously diagnose these failures by using a sophisticated architecture that processes each session as it arrives, accumulates hypotheses, and reasons across the accumulated set to find root causes. The system employs Temporal's durable orchestration capabilities, such as long-running state, durable execution, event-driven coordination, and cross-workflow gating, to build a robust analysis pipeline that identifies failure clusters and generates root cause evidence. This infrastructure enables Kelet to diagnose thousands of sessions daily without human intervention, leveraging Temporal's capabilities to focus on solving AI quality issues rather than building coordination infrastructure from scratch.