EvoX preprint reports higher agent task success through runtime controls

A preprint introduces EvoX, a Rust-based runtime for AI agents, and reports a 56.2% success rate across 691 common task-model pairs. That was the highest numerical result among six agent harnesses, approximately 3.0 percentage points above the average of the other five. The runtime combines guidance to models with deterministic controls for carrying out actions, continuing interrupted execution and grounding completion claims in evidence. Evaluation covered five benchmarks and six foundation models. Gains depended on the model: advantages were +10.7 points on GLM-5.3 and +7.9 points on Kimi K3, with no memory, evolution or extensions used.