Multi-Agent LLM Routing with Ant Colony Optimization: Quality at a Fraction of the Cost

ArXiv: 2603.12933 — Published 2026-03-13

---

The Problem

Running everything through your best (most expensive) LLM is a tax you pay on every single call — even the ones that don't need it.

In a multi-agent system, you have agents with different capability profiles and wildly different costs. The question is: who handles what?

What the Paper Does

Researchers propose using Ant Colony Optimization (ACO) — the same algorithm inspired by how ants find the shortest path to food — to route tasks to the right LLM agent dynamically.

The routing system learns over time which agents perform best on which types of tasks, building a routing map that optimizes the quality-to-cost tradeoff across the whole system.

Key results:

• Maintained benchmark performance comparable to routing everything to the best model
• Significant cost reduction by routing simpler tasks to smaller, cheaper agents
• Interpretable routing decisions — you can see why a task went where it did

Why This Matters for Builders

Most production AI systems today are either:

1. Single model — simple but expensive and bottlenecked
2. Hand-crafted routing — brittle, hard to maintain, doesn't adapt

ACO-based routing is a middle path: learned, adaptive, and explainable. As your agent pool grows and your usage patterns evolve, the routing improves automatically.

This is the kind of infrastructure thinking that separates AI toys from AI products.

Builder Takeaway

If you're running multi-agent pipelines, you should be measuring cost-per-task per agent. The gap between your cheapest and most expensive agents is probably larger than you think — and most tasks don't need the expensive one.

ACO routing is one principled way to exploit that gap systematically.

---

Source: Xudong Wang, Chaoning Zhang, Jiaquan Zhang — ArXiv cs.AI, March 2026