FreshmanD

Facts: - +30% accuracy on IT21HMDB01-B2 - +25% accuracy on H800 - Models already wired into ops pipelines

What’s different: Agents autonomously design and evolve ML models (not manual tuning, not classic AutoML)

Status: Running in production

We built LoongFlow after struggling with prompt-heavy agent systems that didn't improve across runs.

Instead of tuning prompts, we structure agent behavior as an iterative plan–execute–summarize loop, where failures are explicitly summarized and reused in later planning.

The repo includes runnable examples and evolution logs. Happy to answer implementation or design questions.

We’re building LoongFlow, a cutting-edge open-source evolutionary agent framework that integrates large language model reasoning into evolutionary search. If you’re passionate about autonomous agents, AI optimization, or open-source collaboration, we'd love your help in developing new agents, improving algorithms, and expanding benchmarks. Check out the repo here: GitHub Link. Looking for contributors in:

Algorithm development

Benchmarking

Documentation & tutorials

Memory systems & LLM integration Join us and help shape the future of AI!

Each task is unique, unless we provide share memory for them. It means when you start a task, it will run the full evolutionary process from the start.

Hi HN, we are the team behind LoongFlow. We built this framework to use evolve thinking solve any tasks.

LoongFlow brings Evolutionary Algorithms (EA) into the agent workflow. It evolves taskss over generations (via selection, crossover, and mutation) to maximize performance.

Key features:

General-Evolve: Good at Algorithm task.

ML-Evolve: Specialized for machine learning tasks.

Paper: We recently released our paper on arXiv: https://arxiv.org/abs/2512.24077

The repo is fully open source (Python). We'd love to hear your feedback on the architecture and the idea of "breeding" better agents!

Hi HN, we are the team behind LoongFlow.

We built this framework to solve the problem of Agent brittleness—where standard ReAct agents often get stuck or fail when prompts aren't perfectly hand-tuned.

Instead of manual prompt engineering, LoongFlow brings Evolutionary Algorithms (EA) into the agent workflow. It treats prompts and logic as "populations" that evolve over generations (via selection, crossover, and mutation) to maximize performance.

Key features:

General-Evolve: Auto-optimizes prompts and code logic.

ML-Evolve: Specialized for machine learning tasks (AutoML agent).

Paper: We recently released our paper on arXiv: https://arxiv.org/abs/2512.24077

The repo is fully open source (Python). We'd love to hear your feedback on the architecture and the idea of "breeding" better agents!