Read, Route, Patch: Robust Test-Time Evolution for LLM Agents

Large Language Model (LLM)-based agents can improve long- horizon interactive decision-making through test-time evolu- tion, yet existing methods face three information bottlenecks: memory retrieval overemphasizes textual similarity, candidate harnesses are often behaviorally redundant, and whole-harness rewriting may overwrite useful evidence or introduce cross- component inconsistency. We introduce Trident, which improves experience utilization, behavioral exploration, and harness evolution through boundary-aware memory reading, diversity-aware candidate routing, and order-aware insertion- based updating, respectively. We evaluate Trident on five benchmarks spanning text games, web interaction, and cod- ing tasks. Across qwen-32B, mistral-small-3.2-24B, and gpt- oss-20B, Trident achieves the best average performance and consistently outperforms non-learning, memory-based, reflection-based, and automated prompt optimization base- lines. Compared with the strongest baseline, APEX, Trident yields relative average gains of 30.7%, 30.8%, and 38.3% on the three backbones, respectively. Ablation and sensitivity stud- ies further show that the three designs provide complementary gains and remain robust to parameter variations.