TRACE-Router Pins a Whole Agentic Task to One Model at Admission and Beats Per-Call Routing by 7.1 Points With 36% Lower Latency on Terminal-Bench

arXiv 2607.22465·medium signalXBlueskyLinkedInCopy link

TRACE-ROUTER (arXiv 2607.22465, v1 July 24 / v2 July 27) argues that per-call LLM routing is structurally mismatched to agentic work: long-horizon workflows only produce a delayed, task-level outcome, so per-call routers cannot attribute feedback to individual decisions. Instead it assigns each task to a backend once at admission using a contextual bandit, pins every subsequent call to that model, and updates on the task's terminal reward jointly weighting accuracy and latency. On tau2-Bench it outperforms latency-matched interpolation between individual models by 7–8 accuracy points; on Terminal-Bench it lands 7.1 accuracy points above the strongest single-model baseline with 36% lower latency, achieving non-dominated Pareto points across three agentic benchmarks.