AgentRouter cuts AI costs by routing tasks to smaller models
AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows
Artificial IntelligenceComputation and LanguageMachine LearningMultiagent Systems
Summary
Large AI systems often waste money by using their most powerful models for every step of a multi-step task, even when simpler models could do some parts just as well. The authors created AgentRouter, a small and fast tool that decides which level of AI model to use at each step. This approach saves about 72% in costs while keeping quality very close to using only the largest model. It works better than previous systems that only looked at one step at a time and ignored how task complexity changes across steps.
What this means in practice
- •For enterprise ai engineers: Optimize compute costs in multi-step AI workflows by dynamically assigning each step to an appropriately sized model tier.
- •For cloud service providers: Reduce infrastructure expenses by integrating step-level model routing to better allocate large and small AI model resources per subtask.
Authors
Rudrendu Kumar Paul, Sourav Nandy
Abstract
Enterprise agentic systems that route every trajectory step to a frontier model waste 60-80% of their inference budget on subtasks that smaller models handle equally well. Existing routing solutions optimize single-turn query assignment but ignore a property unique to agentic workflows: subtask complexity varies widely within a single trajectory. A planning step may require frontier-class reasoning while a subsequent formatting step needs only a 7B model. We formalize step-level model routing as a sequential assignment problem over agent trajectories and propose AgentRouter, a lightweight classifier (12M parameters, <5ms overhead per step on an A100 GPU) that maps each trajectory step to one of four model tiers using five features extractable at routing time. Trained on 50,000 annotated agent trajectory steps spanning planning, coding, research, and data analysis tasks, AgentRouter achieves 72% cost reduction relative to frontier-only baselines, retaining 97.3% of frontier-only quality (less than 3% degradation in end-to-end task completion); per-step routing accuracy reaches 91% on minimal-complexity steps and 85% on efficient-tier steps, with 76-82% on the harder mid-range and frontier tiers. On the same benchmarks, RouteLLM and FrugalGPT (applied per-step) achieve only 31% and 44% cost reduction respectively, because their single-turn training signal misses trajectory-level quality dependencies.