How TCAR Redefines Enterprise Multi‑Agent Routing with Reason‑First Decision Making

Problem Statement

Enterprise multi‑agent systems rely on a router to dispatch tasks to specialized agents. Conventional routers are simple single‑label classifiers, which leads to three critical failures in production environments:

Inability to handle cross‑domain routing where multiple agents may be relevant.

Failure to resolve conflicts when several agents propose overlapping solutions.

Lack of adaptability to newly added agents, requiring costly retraining.

These limitations make the router a "blind" traffic cop that often misroutes incidents.

TCAR Overview

TCAR (Tencent Cloud Andon Router) is an open‑source 4‑billion‑parameter model that replaces direct label prediction with a reason‑then‑select workflow. The model first generates a natural‑language reasoning chain that identifies the problem scope, relevant technology stacks, and the responsibilities of candidate agents. Only after this explicit reasoning does it output a subset of agents to handle the request.

Key Capabilities

Reason‑then‑Select : Produces a transparent reasoning trace, turning the router into an interpretable decision maker.

Group Collaboration : Returns a set of agents instead of a single choice, enabling collaborative conflict resolution.

Expert Consultation : Each selected agent generates an independent answer; a dedicated RefiningAgent merges these answers into a coherent, conflict‑free final response.

Training Methodology

TCAR is trained in two stages:

Supervised Fine‑Tuning (SFT) : The model learns structured reasoning and agent‑set generation using a Slerp‑based fusion technique that blends multiple expert representations.

Reinforcement Learning / DAPO : A reward function optimizes selection correctness. Two metrics are used:

R1 (Precision‑like) : Encourages the selected set to contain only agents that can actually solve the task, penalizing irrelevant candidates.

R2 (Recall‑like) : Rewards inclusion of all essential agents, preventing omission of critical expertise.

Additionally, a length penalty discourages the model from over‑selecting the entire agent pool.

Evaluation

TCAR was benchmarked on five datasets that reflect high‑conflict, cross‑domain scenarios:

CLINC150

HWU64

MINDS14

SGD

Qcloud

Across all benchmarks TCAR consistently outperformed larger mainstream models, achieving higher success rates while maintaining fast inference and low computational cost due to its 4B parameter size. The model excels especially on ambiguous, multi‑agent routing tasks.

Open‑Source Resources

Model checkpoint: https://huggingface.co/tencent/TCAndon-Router

Source code: https://github.com/Tencent/TCAndon-Router

Paper (arXiv): https://arxiv.org/pdf/2601.04544

Illustrations