tcandon-multi-agent-router

name	tcandon-multi-agent-router
title	TCAndon-Router: Adaptive Reasoning Router for Multi-Agent Collaboration
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2601.04544
keywords	["multi-agent","routing","task-assignment","dynamic-adaptation","agent-collaboration"]
description	Route queries to multiple specialized agents dynamically using reasoning-aware routing that generates natural-language justification before predicting candidate agents. Enables seamless addition of new agents without system redesign. Routes aggregate responses from multiple specialists into coherent final answers, supporting enterprise-scale multi-agent systems with overlapping capabilities.

Problem

Multi-agent systems face critical routing challenges:

1. Static Routing Bottleneck: Traditional single-label routing (query → one agent) can't leverage multiple specialists with overlapping skills
2. New Agent Integration: Adding agents requires retraining routers and redesigning routing logic
3. Capability Overlap Conflicts: When multiple agents could handle a query, picking one wastes alternative perspectives
4. Ambiguous Intent: Many queries don't cleanly map to a single agent; forcing 1:1 assignment loses information

Current routers treat query-to-agent mapping as a classification problem, but real-world queries often benefit from multiple specialized perspectives.

Solution

TCAndon-Router (TCAR) introduces Multi-Candidate Reasoning-Aware Routing:

1. Reasoning-First Routing: Generate natural-language reasoning chain explaining why an agent is appropriate before assigning queries
2. Multi-Candidate Assignment: Predict a set of candidate agents rather than single agent
3. Lazy Agent Integration: New agents register themselves; router adapts without retraining
4. Response Refinement: Aggregate responses from multiple agents with a dedicated Refining Agent producing coherent final answer

When to Use

• Enterprise Multi-Agent Systems: Routing queries to teams of specialized agents (customer service, technical support)
• Overlapping Expertise: Domains where multiple agents have relevant but complementary knowledge
• Scalable Agent Networks: Systems that grow from 5 to 100+ agents over time
• High-Confidence Requirements: Critical decisions benefiting from multiple agent perspectives
• Exploratory Agents: Research systems where diverse viewpoints improve answer quality

When NOT to Use

• For single-agent systems (router adds unnecessary overhead)
• In latency-critical applications (multi-agent routing adds response time)
• When computational budget for running multiple agents is unavailable
• For tasks with clear single-agent ownership (no overlap)

Core Concepts

The framework operates on the principle that reasoning improves routing:

1. Interpretable Decisions: Before assigning agents, explain why in natural language