Decision-Model Controllers Increase Agent Costs Without Improving Success

A study using the Jev 1.13 controller found that delegating decision-making to a dedicated model increased agent costs by 19% while maintaining an identical 73.0% pass rate on Terminal-Bench 2.1. The controller was cheaper than a small LLM but less efficient than simple deterministic rules.