Author(s): Praveen Kumar

Originally published on Towards AI.

Google ADK’s LoopAgent Only Stops on One Signal and ADK 2.0 Is Already Replacing It

You wrote a reviewer loop in Google’s Agent Development Kit. A worker drafts, a critic checks the draft, and when the critic decides the work is good enough it returns “approved.” You set max_iterations=10 as a safety net and expected the loop to exit the moment the critic approved. Instead it ran all ten times, burned ten times the tokens, and only stopped because it hit the ceiling.

The article explains that in ADK’s LoopAgent, a sub-agent “finishing” does not stop the loop; the runtime keeps iterating until either the iteration cap is reached or a yielded event sets event.actions.escalate == True. It clarifies that returning values or setting a flag in session state won’t terminate the loop unless you translate that into an escalate signal. The author demonstrates this with three local examples: one where a worker returns without escalating (the loop only stops at max_iterations), another where the critic emits escalate=True at the quality gate (the loop exits early), and a third showing that LoopAgent is deprecated in favor of ADK 2.0’s Workflow runtime. Finally, it provides broader context on ADK 2.0’s deprecation of LoopAgent/SequentialAgent/ParallelAgent and reinforces the takeaway: iterative agent loops must use an explicit escalation contract (or a cap) to avoid wasting tokens, even as migration to Workflow begins.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI