Agent evaluation and long-horizon multimodal benchmarks face a credibility test

RecreationWorld/RecreationBench, GameHorizon, ExplorationBench, Game Arena, AgentWorld, TraceDance, CompoWorld, EmbodiedMemory-Bench, Harness-Zero, Schrödinger's Code Repository, IndicBankBench, LEGO-Anything, Marathoner, VoxMem, and interactive 3D-world auditing show that visual or demo-level success overstates executable reliability. Turbo Harness adds evidence that instance-adaptive orchestration can materially improve coding-agent outcomes, while MINTEval, persistent-calibration work, and context-management proposals expose unresolved memory and confidence failures. Results remain narrow and require independent validation.