The thread is the Workflow: Durable AI agents without changing Agent code

This post details how Manetu's AgentVisor runtime leverages Temporal Workflows to provide durable execution for AI agents without requiring changes to the agent code itself. It maps Agent Protocol concepts (Thread, Run, Checkpoint, Interrupt) to Temporal implementations (Workflow instance, Workflow Update + Activity, Workflow state, Pending Workflow state + signal). Key engineering challenges addressed include managing state for long-lived agent threads within Temporal's history limits (retaining only the latest checkpoint and relying on replay), differentiating between crash-recovery and human-in-the-loop resumes, handling Continue-As-New rollovers synchronously, and resolving worker identity issues by providing explicit build identifiers. The post also highlights the integration of Temporal's Payload Codec for transparent encryption of agent state.