tencent/TCAndon-Router

TCAndon-Router

<p align="center"> <img src="https://raw.githubusercontent.com/Tencent/TCAndon-Router/refs/heads/main/assets/router.png" width="500"/> </p>

<p align="center"> &nbsp;&nbsp; <a href="https://github.com/Tencent/TCAndon-Router">Github</a>&nbsp;&nbsp; | &nbsp;&nbsp; 📑 <a href="https://arxiv.org/pdf/2601.04544">Paper</a> &nbsp;&nbsp; </p>

🌟 Introduction

In multi-agent systems, the ability to select the appropriate agent(s) to handle a user query is a key determinant of overall system performance.

TCAndonRouter is a reasoning-centric multi-intent routing module whose primary role is to perform agent routing in multi-agent systems. Beyond agent routing, TCAndonRouter can be applied to any intent-routing scenario, including agent skill selection.

The main advantages of TCAndonRouter include:

• —Designed specifically for real-world enterprise applications
• —Supports dynamic onboarding of new agents (intents) New agents can be added simply by appending their descriptions, without retraining
• —Provides transparent and interpretable routing decisions, improving explainability, robustness, and cross-domain generalization, and making post-deployment bad-case analysis easier
• —Effectively resolves agent conflicts caused by overlapping responsibilities, leading to higher-quality final responses. When multiple agents are applicable, TCAndonRouter preserves all relevant agents. Each downstream agent generates its own response, and a Refining Agent subsequently merges these outputs into a single final answer

TCAndonRouter is trained using Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (DAPO), and achieves state-of-the-art performance on large-scale, real-world enterprise datasets, including HWU64, MINDS14, SGD, and the Tencent Cloud ITSM dataset(QCloud).

🔧 How to use

Please refer to GitHub for code usage.

from transformers import AutoModelForCausalLM, AutoTokenizer
from prompt import router_prompt
from utils import load_config

tokenizer = AutoTokenizer.from_pretrained("tencent/TCAndon-Router")
model = AutoModelForCausalLM.from_pretrained("tencent/TCAndon-Router", device_map="auto")

agents = load_config('config/hwu64_config.xml')
query = "Can you recommend any pub in mg road"
prompt = router_prompt.format(agents=agents) + 'user:' + query

messages = [{"role": "user", "content": prompt}]
encoding = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=False,
    return_tensors="pt"
)

outputs = model.generate(encoding.to(model.device), max_new_tokens=2048)
output_text = tokenizer.decode(outputs[0])

Generate Agent Descriptions

If you want to use TCAndonRouter on your own dataset, you need to provide agent descriptions. The required format is defined in config/xxx_config.xml. You can generate agent descriptions using an LLM via generateagentdesc.py, or write them manually.

python generate_agent_desc.py --dataset hwu64 --limit 50

🤝 Citation

If you use TCAndonRouter in your work, please cite our paper:

@article{zhao2026TCAndonRouter,
  title={TCAndonRouter: Adaptive Reasoning Router for Multi-Agent Collaboration},
  author={Jiuzhou Zhao, Chunrong Chen, Chenqi Qiao, Lebin Zheng, Minqi Han, Yanchi Liu, Yongzhou Xu, Xiaochuan Xu, Min Zhang},
  journal={arXiv preprint:2601.04544},
  year={2026}
}