EComAgentBench: Benchmarking Shopping Agents on Long-Horizon Tasks with Distributed Hidden Intent

Researchers have introduced EComAgentBench, a benchmark for evaluating shopping agents on long-horizon tasks, and GLM-5.2, a model designed for sustained engineering work.

EComAgentBench is a benchmark of 662 tasks grounded in real Amazon products and reviews, designed to test shopping agents' ability to uncover hidden intent and make decisions over long horizons^[1]. Each task requires agents to uncover hidden intent, verify candidates against attributes and review evidence, and commit to a single product within 100 tool calls. The benchmark uses typed, source-tagged rubrics to grade every task, attributing each failure to a requirement and its source. Even the strongest models attain only 57.1% overall accuracy in the benchmark^[1]. Meanwhile, GLM-5.2 is a long-horizon model built for sustained engineering work, with a 1M-token context that stably sustains long-horizon work. GLM-5.2 trails Opus 4.8 by only 1% on FrontierSWE benchmark, but still has room to grow on SWE-Marathon benchmark, trailing Opus 4.8 by 13%^[2]. GLM-5.2 outperforms both Opus 4.7 and GPT-5.5 on PostTrainBench and is the strongest open-source model on standard coding benchmarks, improving on GLM-5.1 by a wide margin. GLM-5.2 also introduces effort level control, enabling users to balance model capability against task execution speed and computational cost, and its architecture reduces per-token FLOPs by 2.9× at a 1M context length^[2].