BabyAGI Inside embeds a bounded task-planning loop directly into the weights of Qwen3.5-2B. After receiving an objective, the model alone advances the cycle, stores intermediate results, and stops at a compiled halt token.