A-MapReduce preprint reports faster parallel search by AI agents

Researchers report that A-MapReduce, a framework for coordinating AI agents, reduced running time by 45.8% against representative multi-agent baselines in experiments. Their preprint also reports 5.11% - 17.50% average Item F1 improvements against strong baselines using OpenAI o3 or Gemini 2.5 Pro. The framework divides broad retrieval tasks into parallel searches, aggregates their results and uses memory of previous execution to improve task allocation. Experiments covered five agentic benchmarks, with reported state-of-the-art performance on WideSearch and DeepWideSearch. The findings concern benchmark experiments rather than a production deployment.