Agent Reliability Shifts toward Harnesses and Evaluation

New research and deployment signals show that agent behavior depends heavily on orchestration, tools, memory, repair loops, and hardware interfaces rather than the underlying model alone. Zero-WAM reports a 29.5-point gain on unseen simulated tasks using human videos as specifications, while Anthropic’s MHS proposal targets faster device integration; reliability, physical intuition, reproducibility, and manipulation resistance remain unresolved.