Agent Loops vs Agent Graphs: Google Tested 180 Setups and the Graphs Collapsed by 70%

Google Research tested 180 agent configurations across five architectures and found that multi-agent graph variants degraded performance by 39–70% on sequential reasoning tasks while boosting parallelizable tasks by up to 81%. The study shows the choice between agent loops and graphs depends on task properties, not ideology.

Member-only story Agent Loops vs Agent Graphs: Google Tested 180 Setups and the Graphs Collapsed by 70% The loudest debate in AI engineering right now is deceptively simple: should your coding agent be a plain while-loop, or a graph of orchestrated sub-agents? Google Research quietly answered it with data. In a controlled evaluation of 180 agent configurations across five architectures, every multi-agent variant they tested degraded performance by 39–70% on sequential reasoning tasks — while the same graph-style architectures boosted parallelizable tasks by up to 81%. The answer to “loop vs graph” isn’t a religion. It’s a property of your task, and most people are choosing wrong. I’ve spent the past week going through the study, the July framework releases, and the arguments from both camps. Here’s what the numbers actually say, when each side wins, and working code for both so you can test it on your own workload. Why this fight is happening now On July 18, Peter Steinberger — the PSPDFKit founder whose “stop prompting, design loops” line kicked off the loop-engineering wave — posted a single question: “Are we still talking loops or did we shift to graphs yet?” It pulled 514,000 views, 3,300+ likes, and 602 replies in a weekend. The replies split into exactly two camps, and both have famous names attached.