Agent research targets teams, tools, and scientific reliability

New work explores open training pipelines (Axion), reusable multi-agent team structures (EvoTeam), and realistic tool environments (ToolSpace); multi-turn RL findings emphasize recovery and sustained performance. New science-agent evaluations include Science’s Last Exam, with 50+ tasks across nine disciplines, and FrontierPhysics, where Claude Opus 5.5 reportedly scores 27.4%; both offer useful probes, but benchmark methodology and the gap to dependable research remain central questions.