Nace.AI drops Drex Router for agents
Nace.AI has introduced Drex Router, a specialized small decision model designed to select the optimal LLM or SLM and reasoning effort for each step in an agent workflow. The model delivers 87.6% routing accuracy while avoiding the latency and output token overhead of generative models.
Step-level routing is becoming the critical efficiency bottleneck in production agent systems, making fast non-generative decision models far more practical than routing with general-purpose LLMs. By determining model tier and reasoning effort in a single forward pass, Drex Router tackles the compounding costs and delays of multi-step agent loops.
- –Employs calibrated probability classification rather than autoregressive decoding, drastically cutting latency and token consumption during routing decisions.
- –Posts 87.6% routing accuracy against competitor Jev Router's 86.4%, setting a new benchmark for automated model selection.
- –Dynamically assigns reasoning effort per execution step, routing routine tool calls to fast small models while reserving expensive frontier LLMs for complex tasks.
- –Highlights an architectural shift in agent design toward dedicated "System One" decision models acting as pre-flight controllers.
DISCOVERED
1h ago
2026-11-19
PUBLISHED
2h ago
2026-11-19
RELEVANCE
AUTHOR
omarsar0