LoongFlow is an open-source expert-level Agent development framework maintained by baidu-baige and licensed under the Apache License 2.0. Its name is inspired by Wang Yangming's "Enlightenment at Longchang," and it is dedicated to breaking down the barrier between "knowledge" and "action," transforming professional knowledge into AI productivity.
Purpose and Positioning
LoongFlow is oriented toward complex tasks requiring long-horizon reasoning, helping developers quickly build Agents with domain-expert performance. Its core emphasis is not better mutation, but better thinking, reflection, and experience accumulation.
Main Capabilities
- PES thinking paradigm: The core is the Plan-Execute-Summary loop. The planning stage understands the task and constraints, retrieves relevant experience, and designs an execution blueprint; the execution stage conducts structured experiments and validates intermediate results; the summary stage reflects on success and failure, extracts reusable insights, and writes them into structured memory.
- Continuous learning: It adopts multi-structure fused memory, and the Agent synthesizes experience across task iterations, achieving lightweight learning and evolution of "improvement through running" without heavy retraining.
- Evolutionary memory system: It combines the Multi-Island model with MAP-Elites to maintain diversity, uses adaptive Boltzmann selection to balance exploration and exploitation, and leverages a global evolutionary tree memory to support long-horizon experience retrieval and leap-style reasoning.
Usage
LoongFlow requires Python 3.12 or higher. The code can be obtained through the GitHub repository, and it supports a contribution process (Fork, create a feature branch, submit a Pull Request). It is compatible with any OpenAI-compatible API, including commercial services such as OpenAI and Google, as well as local deployments such as vLLM and sglang; you only need to set llm_config in the configuration to point to the corresponding endpoint. The framework provides an interactive Web interface for real-time evolution tracking, including an evolution tree, cross-generation performance tracking, a code diff viewer, and an island distribution map.
Applicable Scenarios
- General-Agent: Flexible skill-driven coding tasks, from simple applications to bug troubleshooting.
- Math-Agent: Efficiently and stably drives mathematical algorithm design and continuous evolution.
- ML-Agent: Autonomously builds machine learning solutions across the full process and continuously evolves to make breakthroughs.
Empirical Results
In 11 challenges in the fields of geometry and algebra, LoongFlow surpassed all known best results and exceeded AlphaEvolve on 7 of those problems, reaching the latest SOTA. It won medals in all 48 Kaggle competitions in MLE-bench, including 26 gold medals. It has also been validated on problems such as mathematical puzzles and MOE load balancing algorithms.
Limitations and Notes
Some descriptions mention that if Gemini 3 Pro is used for the CirclePacking problem, the total cost is about 10 USD, but a complete pricing system is not provided. More details such as specific model compatibility and performance benchmarks are not yet provided or need to be referenced in the official repository documentation.
