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

What this post added

This post details how Kelet AI built a durable agent for diagnosing AI quality failures using Temporal. It describes a four-level Temporal Workflow hierarchy (Session, Signal, Agent Aggregation, Investigate Issue) to handle long-running state, durable execution, event-driven coordination, and cross-workflow gating. The Session Workflow waits for a silence window before analysis and rebuilds state from the database on startup. Signal Workflows run in parallel for enrichment and interrogation. Agent Aggregation Workflows collect hypotheses across sessions, and Investigate Issue Workflows reason over failure clusters to produce root causes and prompt patches. The post also covers a Temporal plugin for integrating Kelet's monitoring capabilities, propagating session attributes via headers for seamless tracing and debugging of Temporal-based agents.