Temporal, Google ADK and LangGraph: human approval tutorial

Melanie Warrick of Temporal explains how to keep an AI agent waiting for human input without blocking unrelated work. Her ice cream delivery demo coordinates fleet, customer and dispatch agents. Google ADK runs the customer and fleet agents, while LangGraph runs dispatch, with Temporal tracking execution state across them.

The talk introduces workers, workflows and activities. Workers execute code, workflows track the process, and activities handle external interactions such as model and tool calls. Warrick describes Temporal as open source software that can be self-hosted, with a managed service as another option.

Human approval uses a wait condition and a signal. The affected workflow pauses until input arrives while other work continues. Warrick also explains how timeouts can limit the wait. The code walkthrough covers wrapping ADK models and tools, configuring its worker plugin, and connecting graph interrupts to approval responses.

In the recovery demonstration, Warrick kills a worker while an order awaits approval, submits approval while the worker is down, then restarts it. She explains that Temporal queues the event and replays execution history to recover state. This demonstrates the example's recovery path rather than establishing reliability for every deployment. She closes by weighing the cost of an incorrect autonomous decision against alert fatigue from frequent approval requests.