A hands-on exploration of building a durable multi-agent delivery fleet using Google ADK and LangGraph, orchestrated by Temporal. The post demonstrates three deployment modes: all-ADK, all-LangGraph, and a cross-framework setup where ADK handles parallel agent assessment and LangGraph handles dispatch with human-in-the-loop approval. Two human-in-the-loop patterns are covered: an operator interrupting an agent mid-delivery via Temporal Signals, and an agent proactively asking a human for a decision using LangGraph's interrupt() suspended by Temporal's durable wait. Key engineering lessons include separating Task Queues to prevent model inference from starving other activities, using Continue-as-New for long-running driver loops, and why durable execution (not in-memory state) is essential when humans can take hours or days to respond. The post also clarifies the distinction between LangGraph's checkpointer-based durability (framework-local) and Temporal's system-wide durable execution.
