Deep-search agents are commonly trained via repeated interaction with real-world search APIs and browser environments. While realistic, this paradigm is expensive, difficult to reproduce, and sensitive to external drift. Existing simulator-based alternatives reduce dependence on live retrieval, but they largely inherit task distributions from pre-existing QA datasets, limiting their alignment with the behaviors required for deep search. We introduce SynSearch, a framework for training deep-search agents entirely within a synthetic search–browse environment. Our core idea is to jointly design the environment and the training-task distribution, enabling the synthesis of environment-aligned tasks and reinforcement learning of compact language models in multi-turn search settings. By grounding task validity in executable environment trajectories rather than surface-level plausibility alone, SynSearch unifies task generation, trajectory verification, and policy optimization within a low-marginal-cost and reproducible framework. Experiments on multi-hop QA and deep-search benchmarks show that SynSearch consistently improves both answer accuracy and search behavior, with especially large gains on deep-search tasks. Moreover, SynSearch offers low marginal cost, reproducibility, controllability, and strong performance under real-search evaluation, highlighting synthetic environments as an effective substrate for training deep-search agents.

Open-web deep research is limited not only by how long an agent searches, but also by how it organizes evidence into independent, verifiable branches. Existing systems often scale single-path ReAct trajectories or long-context search, which can deepen one hypothesis while leaving alternative entities, constraints, or falsifying evidence underexplored. We propose P-Search, a parallel plan–act–verify framework that first builds a grounded background state, then decomposes the task into independently executable evidence branches, runs branch-specific search agents in parallel, and iteratively verifies unresolved constraints before final synthesis. For training, we avoid end-to-end credit assignment over a full multi-stage rollout. Instead, P-Search-React learns reusable grounded search-report behavior from raw-question search instances and is reused for branch execution and targeted supplementary search, while P-Search-Plan learns schema-constrained routing, decomposition, and verification decisions. We supervise this structure with star-shaped subgraph data and role-specific GRPO rewards. Experiments show that instantiating P-Search with open-source backbones such as GLM-5 and Tongyi-DR-30B improves their reported baselines on most open-web benchmarks, while the trained 8B P-Search-React + P-Search-Plan system reaches 59.2% on GAIA and 70.0% on Xbench-DeepSearch. Ablations confirm the complementary roles of Plan, ReAct, and Verify, as well as the transferability of the trained planner across ReAct backbones.

Modern industrial applications increasingly demand language models that act as agents, capable of multi-step reasoning and tool use in real-world settings. These tasks are typically performed under strict cost and latency constraints, making small agentic models highly desirable. In this paper, we introduce the AgenticQwen family of models, trained via multi-round reinforcement learning (RL) on synthetic data and a limited amount of open-source data. Our training framework combines reasoning RL and agentic RL with dual data flywheels that automatically generate increasingly challenging tasks. The reasoning flywheel increases task difficulty by learning from errors, while the agentic flywheel expands linear workflows into multi-branch behavior trees that better reflect the decision complexity of real-world applications. We validate AgenticQwen on public benchmarks and in an industrial agent system. The models achieve strong performance on multiple agentic benchmarks, and in our industrial agent system, close the gap with much larger models on search and data analysis tasks.