Gemini-3.8-Flash leads in SWECraft's software-agent simulations

SWECraft's researchers say they ran more than 10,000 hours of agent work across three simulated software worlds. Agents maintained interdependent Python packages, and their work was measured by speedups in hidden dependent packages. The team reports Gemini-3.8-Flash led its tests, while GPT-6-Astra scored worst among the frontier models tested; some agents stopped looking for work prematurely. The researchers also found that communication helped some models discover tasks, while clear code ownership reduced duplicated effort.