Fourteen papers tracking how agent eval has hardened year over year. AgentBench, WebArena, SWE-bench, and GAIA established the four corners: multi-environment, browser, code, and general assistant. AgentBoard reframed scoring as progress rate over subgoals. OSWorld and tau-bench made the harness real desktops and real users, and both have moved on since: OSWorld 2.0 (announced 2026-06-26) swaps the 369 original tasks for 108 workflows that take a skilled human 1.6 hours at the median, and the tau-bench repo now ships tau3-bench with voice and a banking knowledge domain, where results from before its v1.0.1 grading fix (July 2026) are not comparable. MLE-bench, SWE-Lancer, BrowseComp, and Terminal-Bench pushed task realism further. Terminal-Bench 2.0 had 89 tasks; the project has since reached 4.0, which removed eight tasks (saturated, refusal-prone, publicly solved, or flawed), fixed 19, and set a flat 8-hour agent timeout, so pin the version when you report a number (checked as of October 2026). AgentDojo and Cybench turn the lens onto security. ScienceAgentBench and SWE-bench Multimodal (in the methodology theme) extend the same instincts to new domains.