A new research paper describes an AI search agent that stops trying to plan, search, and write all with one brain.
Most "deep search" agents today use a single model to juggle everything: figuring out what to look up, digging through search results, and writing the final answer. Researchers behind IterSynth argue that setup breaks down on long, complicated queries, because one policy overloaded with three jobs does all of them worse, and the growing pile of search history buries the useful bits in noise. Their fix splits the work into two roles: a Planner that decides what information is still missing, and a Synthesizer that folds new evidence into a running summary instead of an ever-growing transcript. They also built a training method, Role-Decoupled Policy Optimization, that scores each role separately so credit for a good or bad outcome lands on the part that earned it.
The payoff is a smaller model punching above its weight. An 8-billion-parameter version of IterSynth scored 50.7 on average across five long-horizon search benchmarks, including BrowseComp, beating the best prior agent of similar size by 4.2 percent. The same planner-synthesizer split also boosted bigger proprietary models with no retraining at all.
It is a small paper, but it points at a real bottleneck: as agents are asked to do more autonomous research, cramming every job into one prompt loop is starting to show its limits. Splitting roles and compressing memory into a running summary looks like a cheap, generalizable fix rather than a one-off trick, which is why the zero-shot gains on frontier models matter as much as the benchmark score.