Automating scientific discovery in complex, experiment-driven domains requires more than iterative mutation of programs; it demands structured hypothesis management, environment interaction, and principled reflection. We present OR-Agent, a configurable multi-agent research framework designed for automated exploration in rich experimental environments. OR-Agent organizes research as a structured tree-based workflow that explicitly models branching hypothesis generation and systematic backtracking, enabling controlled management of research trajectories beyond simple mutation-crossover loops. At its core, we introduce an evolutionary-systematic ideation mechanism that unifies evolutionary selection of research starting points, comprehensive research plan generation, and coordinated exploration within a research tree. We further propose a hierarchical optimization-inspired reflection system: short-term experimental reflection operates as a form of verbal gradient providing immediate corrective signals; long-term reflection accumulates cross-experiment insights as verbal momentum; and memory compression serves as a regularization mechanism analogous to weight decay, preserving essential signals while mitigating drift. Together, these components form a principled architecture governing research dynamics. We conduct extensive experiments across classical combinatorial optimization benchmarks-including traveling salesman, capacitated vehicle routing, bin packing, orienteering, and multiple knapsack problems-as well as simulation-based cooperative driving scenarios. Results demonstrate that OR-Agent outperforms strong evolutionary baselines while providing a general, extensible, and inspectable framework for AI-assisted scientific discovery. OR-Agent source code and experiments data are publicly available at https://github.com/qiliuchn/OR-Agent.

翻译：在复杂、实验驱动的领域中实现科学发现自动化，不仅需要程序的迭代变异，更要求结构化的假设管理、环境交互以及原则性的反思。我们提出OR-Agent，一个可配置的多智能体研究框架，专为在丰富的实验环境中进行自动化探索而设计。OR-Agent将研究组织为结构化的树状工作流，显式建模分支假设生成与系统回溯，从而实现对研究轨迹的受控管理，超越了简单的变异-交叉循环。其核心是一种进化-系统性构思机制，该机制将研究起点的进化选择、全面研究计划的生成以及研究树内的协调探索统一起来。我们进一步提出一种受分层优化启发的反思系统：短期实验反思作为一种“语言梯度”提供即时纠正信号；长期反思则积累跨实验见解作为“语言动量”；而记忆压缩则充当类似于权重衰减的正则化机制，在保留关键信号的同时减轻漂移。这些组件共同构成了一种管理研究动态的原则性架构。我们在经典组合优化基准测试（包括旅行商问题、带容量约束的车辆路径问题、装箱问题、定向问题以及多重背包问题）以及基于模拟的协同驾驶场景中进行了大量实验。结果表明，OR-Agent在超越强进化基线的同时，为AI辅助的科学发现提供了一个通用、可扩展且可检视的框架。OR-Agent的源代码与实验数据已在 https://github.com/qiliuchn/OR-Agent 公开提供。