Agent research pushes toward cheaper long-horizon and robotic deployment

EdgeWise reports cost cuts of up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1 through selective cloud delegation, while claiming quality comparable to or better than full-LLM execution. NVIDIA’s open-source ASPIRE framework targets real-robot skill discovery and self-improvement, with striking but insufficiently detailed benchmark claims. Independent replication and evidence across settings will determine how significant these results are.