Show HN: Synthesize complex agent training data with just a few lines of code (github.com)

A groundbreaking tool, AgentFlow, has been unveiled as the first unified agent data synthesis framework designed to streamline the generation of high-quality training and evaluation datasets across diverse agent environments. This all-in-one platform supports multiple domains—including RAG, MM-Doc, and Deep Research—allowing researchers to synthesize intricate agent trajectories, reasoning paths, and tool interactions with minimal coding effort. The framework not only facilitates data synthesis but also enables joint training of models on assorted trajectories, promoting the development of robust, industrial-grade agentic models capable of functioning seamlessly across a variety of tasks. The significance of AgentFlow lies in its ability to provide a comprehensive abstraction layer for diverse environments, thereby enhancing the research landscape in AI/ML. The innovative three-stage data synthesis pipeline—comprising trajectory sampling, selection, and QA synthesis—ensures the quality of data generated. Furthermore, it offers a suite of rigorous benchmarks crafted to evaluate emerging agent capabilities and address real-world challenges typically overlooked in existing assessments. As a result, AgentFlow promises to accelerate advancements in agent research, making it easier for developers to create sophisticated artificial intelligence systems.