Agent harness optimization research accelerates

New benchmarks and studies focus on improving agent harnesses for practical deployment. Evo-Bench isolates harness improvement from base model strength, showing top models gain 16.6 points. HarnessOpt-Bench tests LLMs' ability to optimize prompts, tools, and control flow. Studies challenge self-reflection loops, finding no reliable win in 36 comparisons. Cost-aware comparison with repeated sampling is actionable. Simple harnesses like Pi perform best. These developments directly impact self-hosted agent efficiency.