KwaiAgents is an open-source LLM agent system from Kuaishou’s KwaiKEG research group, designed for generalized information-seeking tasks. Hosted on GitHub with roughly 1,200 stars, it shows how large language models can plan, search and reason their way through questions that require gathering facts from multiple sources rather than answering from memory alone.
KwaiAgents was released alongside research on agent tuning, including the KwaiAgent-Qwen models and the KAgentBench evaluation benchmark. The system orchestrates a set of tools — web search, browsing and other information-gathering utilities — under the control of an LLM that breaks a question into sub-tasks, calls tools, and synthesizes a final answer from what it finds. Crucially, the project also contributes training data and fine-tuned open models intended to close the gap between proprietary agents and open-source ones, giving researchers a complete reference stack: data, models, benchmark and a working agent framework. It supports local LLMs as well as API-backed models, and the repository documents how to reproduce the paper’s experiments and run the agents yourself.
The project comes from Kuaishou’s knowledge-engineering group (KwaiKEG) and was presented as a full research package rather than a demo: the agent framework, the tuning dataset and the benchmark were all released together so results can be examined and extended by the community.
KwaiAgents is free and open source; the code and research artifacts are published on GitHub at no cost. Running it may involve your own compute or third-party LLM API charges depending on the models you plug in.
AI researchers and engineers studying autonomous agents, agent fine-tuning and tool use, who want an open reference implementation with data, models and benchmarks included.