Anthropic adds evals to Agent Skills

This is the missing layer between “cool demo” and production-grade agent behavior.

• –Anthropic is explicitly framing Skills as software that should be evaluated against real failure cases, not just hand-waved with better prompts
• –The PDF form example is the key tell: Skills now have a path to catch brittle edge cases before they quietly break user workflows
• –The docs pair well with this launch by pushing concise instructions, validation loops, and model-specific testing across Haiku, Sonnet, and Opus
• –For AI teams, this makes Claude’s agent stack feel more like CI for workflows: define expected behavior, measure it, refine, repeat