How Kelet AI built a durable agent that debugs durable agents.

This post details the architecture and implementation of Kelet AI's durable agent for diagnosing AI quality failures, built on Temporal. It introduces a four-level Workflow hierarchy (Session, Signal, Agent Aggregation, Investigate Issue) to address the challenge of identifying root causes across multiple AI agent sessions. Key Temporal primitives utilized include long-running state, durable execution, event-driven coordination via Signals, and cross-workflow gating. The post also describes a Temporal plugin for integrating Kelet's monitoring, which propagates session attributes via Temporal headers for deterministic replay and survival of worker restarts and `continue_as_new`.