The article proposes Harness-Native agentic routing, a step-level paradigm that selects models based on the full execution harness state rather than optimizing single-turn cost-quality trade-offs. This approach treats every routing decision as a structured data record to form a harness-native data flywheel, where execution traces train better routers and models.
- The system is instantiated in OpenSquilla using a four-layer routing stack, an open LightGBM cold-start ranker, and a staged router-model path.
- It enables selection of either a single best-fit model for cost-effective execution or multiple complementary models for ensemble-style accuracy improvement.
- The report evaluates singleton and multi-model routing on agentic benchmarks including DRACO and PinchBench.
The authors argue that agentic routing serves as a data engine for agent-native training, improving cost-quality trade-offs by generating more traces under the same budget.