Research
Rogue trains a router that cuts coding-agent cost by 15%
The router reads the task before the agent starts and picks the reasoning effort. On Terminal-Bench 4.0 it matched high effort at 15% lower cost.
Rogue published “Your Agent Is Overthinking”, research on how much reasoning a coding agent should spend on each task. The router is a small model that reads the request before the agent starts and sets the effort.
Trained on DeepSWE and tested on Terminal-Bench 4.0, it matched GPT-6 Astra's high-effort pass rate at 15% lower cost and finished tasks about 7% faster. Running everything at max effort cost 69% more than the router for a score only 0.3% higher.
The recipe applies to any model: it needs fewer than a thousand trajectories and trains in under an hour. Next, Rogue will bring it to its products.
