Agent Infrastructure, Evaluation, Harness Design, and Multi-Model Economics

QA Bench’s browser-agent results did not test Manus, while SealedBench’s proposed commitment protocol and documented flawed runs highlight benchmark leakage and evaluator discretion. A new harness catalog emphasizes how tools, permissions, memory, and recovery shape agent performance, though its rankings need scrutiny. MAPLE’s low-VRAM Gemma 4 MoE demonstrations remain feasibility signals with limited quality checks.