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

1 Introduction

• Task routing assigns queries to domain experts, but overlapping responsibilities and multi-intent queries can make single-label decisions brittle and reduce reliability.Figure 1 contrasts performance routing, which balances latency and cost across model sizes, with task routing, which assigns domain-specific experts to improve accuracy.Source: §1 · pp. 1–2, Figure 1 · p. 2
• TCAR generates natural-language reasoning and selects all potentially relevant expert agents rather than making a single-label routing decision.The paper presents this reasoning-centric approach as supporting interpretable decisions about when and why multiple agents may apply.Source: §1 · p. 2
• Selected experts produce domain-specific responses, which a Refining Agent aggregates and reconciles into a coherent final answer.The pipeline is presented as a way to integrate complementary domain expertise and mitigate agent conflicts.Source: §1 · p. 2
• Across public benchmarks and real-world cloud scenarios, combining multi-agent routing with refinement improves accuracy, reduces routing conflicts, and strengthens robustness on ambiguous or cross-domain queries.Source: §1 · p. 2
• The authors release TCAR and associated resources to support research on interpretable multi-agent routing and reasoning-driven collaboration.Source: §1 · p. 2

2 Related Work

• Performance-based routers select among differently sized models to balance cost and latency, while task-based routers assign queries to domain experts for accuracy.Source: §2 · p. 2
• Earlier task routers used fixed single-label classification, while later similarity-based methods sought to support adding agents without retraining.Examples include KNN similarity [21] and query–agent dual-tower or fusion-based cross-encoder models.Source: §2 · p. 3
• More recent LLM-based routers still generally select one agent and lack explicit structured reasoning for cases where multiple agents are relevant.Examples include ArchRouter [5], AgentRouter [23], and Confidence-Driven LLM Router [24].Source: §2 · p. 3
• Collaboration frameworks support joint task execution, but typically lack systematic selection of relevant experts and coherent integration of their outputs.Examples include AutoGen, CAMEL, and CrewAI.Source: §2 · p. 3

3.1 Problem Formulation

• The router prompt concatenates its instruction, the user query, and descriptions of all available agents, and outputs both reasoning and a selected agent set.Source: §3.1 · p. 3

3.2 Dynamic Agent Addition

• Adding an agent requires appending its natural-language description to the agent set, without modifying or retraining the router.Source: §3.2 · p. 3

3.3 Multi-Agent Conflicts

• TCAR replaces single-label routing with a reason–then–select process that outputs a reasoning chain and a subset of potentially relevant agents.The reasoning examines query causes, technical stack, and agent responsibilities, supporting interpretation of selection decisions.Source: §3.3 · p. 4
• Selected agents answer in parallel, and a Refining Agent filters, aligns, and fuses their responses into one final answer.Unlike the solo-agent flow, the refining flow routes to candidate agents and aggregates their outputs.Source: §3.3 · p. 4, Figure 2 · p. 5

3.4 Training TCAR

• TCAR first uses supervised fine-tuning to teach reason–then–select outputs, then applies DAPO reinforcement learning to improve selection and reasoning.Source: §3.4 · pp. 4–5
• SFT may lead to overfitting annotation templates or mechanical reasoning, motivating further improvement to reasoning quality and agent-selection accuracy.Source: §3.4 · p. 5
• The reinforcement-learning reward combines predicted-set overlap with a coverage reward that applies only when the predicted set contains the full gold set.The overlap reward measures the proportion of predicted agents that are correct.Source: §3.4 · p. 5
• The final reward weights precision-like correctness against recall-like coverage and penalizes predictions exceeding the gold-set size.The length penalty is intended to prevent overly long outputs or endlessly listing agents.Source: §3.4 · pp. 5–6

4.1 Experimental Setup

• Evaluation spans four public intent datasets and internal QCloud, using accuracy for single-agent tasks, F1 for multi-agent tasks, and end-to-end task success for overall responses.QCloud covers cloud domains and is described as having stronger domain coupling, greater intent ambiguity, and more frequent agent conflicts than public datasets.Source: §4.1 · p. 6

4.2 Main Results

• TCAR, with 4B parameters, achieves state-of-the-art performance on all datasets except CLINC150 and surpasses leading general-purpose LLMs in QCloud F1.Table 1 covers five datasets; the authors attribute the CLINC150 exception to its 150 intent categories and prompts averaging 18k tokens.Source: Table 1 · p. 7, §4.2 · p. 7
• For troubleshooting queries, refinement of multiple agents’ responses produces more comprehensive and reliable answers than relying on an individual agent.Figure 3 reports a significantly higher human-preference win rate for refinement, particularly on troubleshooting queries.Source: §4.2 · p. 7, Figure 3 · p. 8

4.3 Ablation Studies

• Adding explicit reasoning consistently improves routing results across public datasets and QCloud, and the authors report improved generalization.The ablation compares direct agent-set prediction with generating a reasoning chain before predicting the agent set.Source: §4.3 · p. 7
• On QCloud, RL improves recall while maintaining high precision, mitigating the tendency of SFT-only models to output a single agent.The authors report consistent and significant improvements from RL over SFT-only training.Source: §4.3 · p. 8
• Slerp initialization yields similar final metrics but higher entropy and early exploration; the authors associate it with better generalization and less entropy collapse.The Slerp-merged model also has higher early pass@k, while the reward curves are similar to those from SFT initialization.Source: §4.3 · pp. 8–9, Figure 4 · p. 9

5.1 Error Analysis

• TCAR’s routing errors stem from incomplete queries, gaps in specialized-domain understanding, and mismatches between reasoning and final predictions.The authors identify reasoning-to-prediction alignment as an open problem.Source: §5.1 · p. 9

5.2 Multi-Agent Collaboration Analysis

• In Tencent Cloud troubleshooting, the Refining Agent uses RAG to compare and filter candidate answers, reconciling conflicts and combining complementary partial solutions.The authors note that a single agent is typically insufficient for these scenarios.Source: §5.2 · p. 10

5.3 Efficiency and Cost Analysis

• 1.37 agents are routed on average; reasoning is under 100 tokens, GPU latency is less than 1 second, and downstream execution adds 0.37 agent calls on average.The authors report that the low average number of selected agents keeps downstream collaboration overhead small.Source: §5.3 · p. 10

Agent description quality

• Routing accuracy depends on manually written agent descriptions that adequately represent each agent’s responsibilities.Brief, ambiguous, or incomplete descriptions may misdirect the reasoning chain; reinforcement learning improves robustness to description noise but does not remove this dependence.Source: §6 · p. 10

Long-tail knowledge and domain transfer

• Low-frequency or highly specialized scenarios may remain unstable when training samples are insufficient.Substantial instruction-format changes may also degrade instruction following; the authors suggest fine-tuning on domain-specific private data in that case.Source: §6 · p. 10

Scope

• The authors state that these limitations do not undermine TCAR’s effectiveness in mainstream routing tasks or real enterprise deployments.They identify extreme or specialized conditions as directions for future robustness and generalization work.Source: §6 · p. 10

Contribution

• TCAR combines natural-language reasoning and subset prediction with a collaborative execution mechanism and Refining Agent that consolidates agent outputs.The framework targets agent conflicts, cross-domain intent ambiguity, and limited interpretability.Source: §7 · p. 10

Results

• Experiments on public datasets and large-scale Tencent Cloud ITSM data show substantial improvements in task accuracy, recall in conflict-prone scenarios, and final answer quality from collaboration.The authors also report that ablations validate reasoning chains, reinforcement learning, and SLERP-based model fusion.Source: §7 · pp. 10–11

Analysis and future work

• The authors identify reasoning-chain structure, agent-description quality, and sparse long-tail domain knowledge as factors influencing performance.Future work will explore structured reasoning constraints, more efficient collaboration protocols, and lower-cost expansion for emerging domains.Source: §7 · p. 11

Router prompt

• The Router Agent selects agents for a user issue based on chat history, with a maximum of three selections.Source: A prompt · p. 15
• The router is instructed to check that agent capabilities fit the problem and to output “oos” when no agent can reasonably handle it.Source: A prompt · p. 15
• The router prompt requires XML output with one or more agent ID lines and a reason for each selection.Source: A prompt · p. 15

Refining Agent prompt

• The Refining Agent compares agent answers, integrates complementary content, and explicitly explains conflicts in its single final response.If all answers fail to provide useful information or refuse, the prompt instructs it to respond with a polite refusal.Source: A prompt · p. 15

• The bad-case categories include incomplete problem descriptions and highly domain-specific scenarios.Source: B Badcase · p. 15
• For a slow-webpage query, TCAR assigns diagnosis to the Network, CDN, and EdgeOne agents.The reasoning cites network-path and acceleration-layer diagnosis as relevant.Source: B Badcase · p. 15
• For a CKafka configuration-change query, TCAR instead selects the RocketMQ Message Queue Agent, whose listed scope includes configuration changes.Source: B Badcase · p. 16