AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows

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 miss a property unique to agentic workflows: subtask complexity varies widely within a single trajectory, so a planning step may need frontier-class reasoning while a later formatting step needs only a 7B model.

AgentRouter formalizes step-level model routing as a sequential assignment problem over agent trajectories. It is a lightweight 12M-parameter classifier with less than 5ms overhead per step on an A100 GPU, mapping each trajectory step to one of four model tiers using five features extractable at routing time. It was trained on 50,000 annotated agent trajectory steps spanning planning, coding, research, and data analysis tasks.

The system achieves a 72% cost reduction relative to frontier-only baselines while retaining 97.3% of frontier-only quality, amounting to 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.