Instead of hand-designing which AI agents talk to each other and what each one does, can one router decide?
Can one routing framework cover topology and role choices in multi-agent systems?
This explores whether the choices about how agents are wired together (who talks to whom) and what job each agent does (its role) can be made by one router, instead of being designed separately by hand.
This explores whether one router can make both the wiring choice (who talks to whom) and the role choice (what each agent does), instead of having them designed separately. The corpus says yes, and its main example goes further than topology and roles.
The direct answer is MasRouter. It treats routing in a multi-agent system as four decisions made together: the collaboration topology, how many agents to use, which role each agent gets, and which LLM powers each one. A cascaded controller makes these choices in sequence, so each one is conditioned on the ones before it What decisions must multi-agent routing systems optimize simultaneously?. It beat single-model routing by 3.51% accuracy while cutting HumanEval cost by 49%. The reason to unify is that the decisions depend on each other. A good role set depends on the topology, and the right model depends on the role, so picking them one at a time bakes in mismatches.
The cost savings may say something about where the gains come from. Small language models handle most repetitive agent subtasks at 10–30× lower cost, which makes per-agent model assignment worth optimizing Can small language models handle most agent tasks?. Separately, one study found that about 80% of multi-agent performance variance comes from token budget rather than coordination cleverness How does test-time scaling work at the agent level?. Put together, one reading is that a unified router is partly a budget allocator that decides where tokens and model strength get spent. That reading is an inference and isn't tested directly.
There is also evidence that these decisions may not all belong in one router. In a 25,000-task experiment, the winning setup fixed the structure externally (a sequential order) but let each agent pick its own role. It beat centralized control by 14% and fully autonomous systems by 44%. Agents invented specialist roles and abstained when they weren't competent Do self-organizing agent teams outperform rigid hierarchies?. MasRouter centralizes topology and roles together, while this result suggests splitting them, with topology imposed and roles emerging. The corpus doesn't compare the two head to head, so which is better is an open question.
The same routing idea shows up at other scales, which suggests a framework that spans more than topology and roles. Skill routing becomes decompose, retrieve, then compose, because real tasks need several skills assembled into a plan How should agents route across thousands of skills?. Capability vectors let agents be discovered by semantic match under policy and budget constraints, with no manual wiring Can semantic capability vectors replace manual agent routing?. A deployed router also logs trajectories and outcomes that can be turned into training data, so the router can improve itself Can a routing harness generate its own training data automatically?. The pressure to get structure right is real. Single agents hit organizational limits that more capability can't fix Do single agents always hit organizational limits?, and coordination degrades as groups grow Why do multi-agent systems fail to coordinate at scale?.
Sources 9 notes
MasRouter shows that routing in multi-agent systems must jointly optimize collaboration topology, agent count, role allocation, and per-agent LLM assignment through a cascaded controller. This unified approach surpasses single-model routing by 3.51% accuracy while cutting HumanEval costs by 49%.
SLMs handle the repetitive, well-defined language tasks that constitute most agent work at 10–30× lower cost than LLMs, making heterogeneous architectures (SLMs by default, LLMs selective) the economically rational design pattern.
Research shows 80% of multi-agent performance variance comes from token budget, not coordination intelligence. LatentMAS and shared-KV-cache approaches offer ways to decouple performance gains from token costs.
A 25,000-task experiment across 8 models and multiple agent counts showed that sequential protocols with external ordering but internal role selection outperform centralized systems by 14% and fully autonomous systems by 44%. Agents spontaneously invented specialized roles and self-abstained when incompetent.
Real agent tasks demand composing multiple skills rather than selecting one. The routing problem becomes decompose-retrieve-compose: break queries into atomic sub-tasks, retrieve skills per sub-task, and assemble them into dependency-aware plans.
Show all 9 sources
Versioned Capability Vectors embedded in HNSW indices couple semantic matching with policy and budget constraints, making capability discovery a first-class operation that scales sub-linearly as agent heterogeneity increases.
A deployed routing system records execution trajectories, capability demand estimates, and outcome data that can be converted into labeled training examples for fine-tuning and distillation, turning the harness into both a serving component and a difficulty labeler.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Towards a Science of Scaling Agent Systems
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- How we built our multi-agent research system
- From Model Scaling to System Scaling: Scaling the Harness in Agentic AI
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- MasRouter: Learning to Route LLMs for Multi-Agent Systems
- AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities