You saw the Shopify results in Section : % faster rendering, % fewer allocations. Here's what the agent did.
It replaced Liquid's regex-based tokenizer with String#byteindex, about % faster for single-byte searches. It cached small integer to_s conversions to avoid Integer#to_s allocations per render. It pre-computed frozen strings for integers through . Each change was small. Stacked across commits from about experiments, they added up to a % speedup on a -year-old Ruby codebase that human engineers had already optimized.