# TRACE-ROUTER: Task-Consistent and Adaptive Online Routing for Agentic AI

Source: [arXiv](https://arxiv.org/abs/2607.22465v1)  
Feed7 permalink: https://feed7.dev/p/2607-22465v1-1g7nw7j  
Published: 2026-07-24T16:29:06.000Z  
Trust: Needs Review (needs_review)

## Why Included

TRACE-Router selects one model per agent task, keeps every call on that backend, and learns from the final outcome. Its benchmarks suggest task-level routing can improve accuracy and latency together.

## Source Summary

TRACE-Router chooses a model when a task arrives, pins every subsequent call to that backend, then updates a contextual-bandit policy from the terminal reward. The reward accounts for accuracy and latency, without explicit task-complexity estimates.

## Practical Implication

For long-running coding agents, route at the workflow boundary rather than treating each model call independently. The reported gains include **7–8 accuracy points on tau2-Bench** and **7.1 accuracy points with 36% lower latency on Terminal-Bench**.

## Agent-Ready Context

TRACE-Router chooses a model when a task arrives, pins every subsequent call to that backend, then updates a contextual-bandit policy from the terminal reward. The reward accounts for accuracy and latency, without explicit task-complexity estimates.

For long-running coding agents, route at the workflow boundary rather than treating each model call independently. The reported gains include **7–8 accuracy points on tau2-Bench** and **7.1 accuracy points with 36% lower latency on Terminal-Bench**.

The evidence covers **three agentic benchmarks** and reports Pareto improvements, but the material does not describe deployment overhead, model-switching constraints, or behavior on real coding workloads.

## Context Map

- Layer: infra
- Domains: coding
- Topics: gateways, agent-reliability, model-selection

## Uncertainty

- The evidence covers **three agentic benchmarks** and reports Pareto improvements, but the material does not describe deployment overhead, model-switching constraints, or behavior on real coding workloads.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.