TCAndon-Router: Adaptive Reasoning Router for Multi-Agent Collaboration

TL;DR

TCAR addresses enterprise multi-agent routing with overlapping domains by reframing routing as multi-agent subset selection guided by explicit reasoning chains. It introduces dynamic agent onboarding via prompts and a Refining Agent that fuses parallel agent outputs into a coherent final answer. The method combines a reasoning-based two-stage training regime (SFT followed by DAPO-based RL) and a reward structure that balances precision and coverage, achieving improved accuracy and recall under conflicts across public and real-world datasets. The work provides an open-source framework to advance interpretable and collaborative multi-agent routing in real-world enterprise environments ($A_q \subseteq A$; $A' = A \cup \{a_{new}\}$).

Abstract

Multi-Agent Systems(MAS) have become a powerful paradigm for building high performance intelligent applications. Within these systems, the router responsible for determining which expert agents should handle a given query plays a crucial role in overall performance. Existing routing strategies generally fall into two categories: performance routing, which balances latency and cost across models of different sizes, and task routing, which assigns queries to domain-specific experts to improve accuracy. In real-world enterprise applications, task routing is more suitable; however, most existing approaches rely on static single-label decisions, which introduce two major limitations: (i) difficulty in seamlessly integrating new agents as business domains expand, and (ii) routing conflicts caused by overlapping agent capabilities, ultimately degrading accuracy and robustness.To address these challenges, we propose TCAndon-Router(TCAR): an adaptive reasoning router for multi-agent collaboration. Unlike traditional routers, TCAR supports dynamic agent onboarding and first generates a natural-language reasoning chain before predicting a set of candidate agents capable of handling the query. In addition, we design a collaborative execution pipeline in which selected agents independently produce responses, which are then aggregated and refined into a single high-quality response by a dedicated Refining Agent.Experiments on public datasets and real enterprise data demonstrate that TCAR significantly improves routing accuracy, reduces routing conflicts, and remains robust in ambiguous scenarios. We have released TCAR at https://huggingface.co/tencent/TCAndon-Router to support future research on explainable and collaborative multi-agent routing.

Figures (4)

• Figure 1: (a) performance router, which balances latency and cost across models of different sizes. (b) task router, which assigns queries to domain-specific experts to improve accuracy.
• Figure 2: (a) Solo Agent: the router outputs a single agent, and that agent directly responds to the user's query.(b) Refining Agent: the router outputs a set of candidate agents, and the refining agent integrates their outputs into a final answer.
• Figure 3: In scenarios where consultation-type and troubleshooting-type queries create conflicts among candidate agents, we compare two strategies: randomly selecting a single agent to respond versus allowing multiple agents to answer followed by aggregation by the Refining Agent. The results show that the Refining Agent achieves a significantly higher human preference win rate, particularly for troubleshooting-type queries.
• Figure 4: Comparing RL initialized from the SFT model and from the Slerp-merged model, we observe that both approaches achieve almost the same reward throughout training. In contrast, the Slerp-initialized model maintains higher entropy, suggesting that it explores a broader solution space.