IterSynth introduces a role‑decoupled iterative synthesis framework for LLM‑based deep‑search agents that separates planning from answer generation. A Planner module proposes sub‑queries or information needs, while a Synthesizer module incorporates retrieved evidence into a persistent summary state that evolves after each search turn. By maintaining the summary as the sole context, the method mitigates role coupling and limits context‑history noise that degrades ReAct‑style agents. Training employs Role‑Decoupled Policy Optimization (RDPO), a reinforcement‑learning scheme that blends terminal outcome rewards with turn‑level rubric scores and computes advantage functions separately for the Planner and Synthesizer, enabling finer credit assignment. Empirical evaluation on five long‑horizon deep‑search benchmarks—including BrowseComp and Xbench‑DS—shows that the 8‑parameter‑billion variant (IterSynth‑8B) attains an average score of 50.7, outperforming the previous best ≤8B agent by 4.2 percentage points. Furthermore, when used as a model‑agnostic prompting strategy, IterSynth yields notable zero‑shot improvements over ReAct and related prompts on proprietary frontier models, demonstrating its generality and effectiveness for complex information‑seeking tasks.