Research questionHow can long-horizon LLM agents learn when to group actions without overcommitting?Long-horizon agents can spend many decisions executing routine sequences, but poorly learned chunk lengths can either waste decisions or carry out too many actions before adapting.