The Human Is an Async API — Melanie Warrick, Temporal
- Source
- youtube.com
- Author
- AI Engineer
- Date
Many agent demos ask a human for approval with a blocking call, which works until the process restarts or the person takes a day to answer. Melanie Warrick of Temporal argued for treating the human as an asynchronous API instead. In her approach the agent runs inside a durable workflow. When it needs approval it waits on a condition, and the human's answer arrives later as a signal. Because the workflow state is persisted, the wait can last minutes or weeks. She showed the pattern with both Google ADK and LangGraph, and demonstrated recovery by killing the worker in the middle of an approval and letting the workflow resume where it stopped. The point for builders is that approval steps are a reliability problem, not just a user interface problem. Any agent that touches money, production systems or customer messages needs a pause that survives deploys and crashes, and durable execution is one established way to get it.
Context
An waiting for human approval needs a way to receive the decision later. Workflow platform Temporal documents this pattern using a wait condition and a signal. The workflow pauses until an approval variable changes; a separately delivered message changes that variable and allows execution to continue.
The conference session applies this pattern to human involvement in agent tasks, including approval of a high-value order. The useful design choice is to represent approval as part of the workflow's state. The person's response can then arrive on a human timescale, while the software has an explicit condition governing its next action.
- Temporal's Melanie Warrick argues durability belongs in the , using Temporal's workers, workflows and activities to track state, shown with Google ADK and LangGraph agents.
- Two Temporal primitives, a wait condition and a signal, let an agent pause for a human and resume when the answer arrives, in both human-to-agent and agent-to-human flows.
- In the live demo she kills the worker mid-approval, then brings it back and the workflow recovers and replays with no lost work.
- On whether a human should be in the loop at all, she weighs the cost of the agent being wrong against alert fatigue.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- human-in-the-loop — Requiring a person's approval at specific points in an automated process, chosen so the irreversible steps are the ones a human sees.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Checking sign-in…
Loading comments…






