Georgia Tech's Topaz Adds Audit Trails to Agentic Model Routing

Model routers inside agentic systems have become the invisible plumbing that decides whether your customer-support subtask goes to a frontier model or a cheap one, and until now those decisions have mostly been silent. A new paper from Georgia Tech, posted to arXiv and accepted as a Spotlight at the ACM CHI 2026 Human-Centered Explainable AI workshop, argues that the silence is itself the problem. Without a rationale, developers cannot tell whether a router is being cleverly frugal or quietly degrading quality by picking a weaker model to save money, what the authors call the difference between "intelligent efficiency" and "latent failures caused by budget-driven model selection."

The authors, Mika Okamoto, Ansel Kaplan Erol, and Mark Riedl, propose a framework they call Topaz. It has three moving parts. The first is skill-based profiling, which synthesizes performance across diverse benchmarks into granular capability profiles for each candidate model. The second is a routing algorithm using budget-based and multi-objective optimization that produces traces of how skill-match scores were weighed against costs. The third is a layer that turns those traces into developer-facing explanations, so a team can read why a task went where it did and iteratively tune the cost-quality tradeoff instead of guessing.

The framing echoes a broader shift in agentic tooling from raw performance to operability. The question is no longer just how good the pipeline is, but whether the team running it can explain and steer it. In the abstract's phrasing, Topaz is meant to help users "understand, trust, and meaningfully steer routed agentic systems," which is the same demand enterprise buyers already make of their MLOps and observability stacks.

A few things to keep grounded. This is a workshop Spotlight rather than a production deployment report, and the arXiv listing that was retrieved does not describe a controlled user study or a head-to-head measurement of cost or quality against an opaque baseline router. The usefulness of natural-language explanations also depends on whether the underlying skill profiles match the tasks a real agent sees, and benchmark-derived skills tend to be slippery outside their distribution.

If routers become a standard tier in agent stacks the way vector stores and evals already have, the teams that ship them with legible decision traces are the ones regulated buyers will actually clear for use.