Agentic harness evolution optimizes an agent's instructions, tools, and control flow by treating task execution experiences as evolving evidence. Although adding more evolution tasks seems beneficial, our initial study uncovered a paradox: performance drops as the task pool grows. Deeper analysis revealed that suboptimal evidence selection during optimization is the root cause, since different evidence mixes lead to vastly different outcomes. To tackle this, we introduce Successive Scale Racing (SSR). SSR adaptively searches for the best evidence combinations each round, builds diverse sets based on task quality and coverage, and races the resulting candidate harnesses under a limited compute budget. Empirically, SSR reaches state‑of‑the‑art results across three benchmarks and six configurations. Notably, when the task pool expands from 40 to 200, SSR’s pass rate rises from 64.3% to 68.3%, while all baselines decline. Its racing mechanism also identifies near‑optimal candidates while cutting validation executions by 60%.
Review: SSR’s fine‑grained evidence selection and efficient racing overcome the scalability bottleneck, offering a practical route for self‑optimizing agentic systems.