How do I know my agent got better between versions and did not just get lucky?

When an agent takes hundreds of sequential steps over hours, traditional A/B evaluation breaks down because an early tool call shapes every subsequent choice, making outcomes path-dependent across runs. Running enough independent trials to get reliable signal costs as much compute as training. The field defaults to proxy metrics such as step success rate and tool call accuracy that demonstrably do not correlate with end-task outcomes on real work. A 2025 audit of 445 published LLM benchmarks documented construct-validity failures at scale: vague task definitions, repurposed short-horizon datasets, and missing statistical tests, all of which become more severe as task horizon grows. Teams building agentic products ship on manual spot-checks because no principled, reproducible evaluation methodology for long-horizon agents exists.