Researchers introduce SHIFT, a method that builds a custom agent harness for each query instead of relying on a single fixed design. A local LLM architect and a learned value function predict the accuracy-versus-cost tradeoff of candidate harnesses, and Monte Carlo tree search uses these predictions to assemble the harness without executing alternatives at inference time. Across 9,193 tasks spanning six benchmarks, SHIFT attained about 80% mean accuracy, outperforming 17 baseline methods spanning prompting, prompt optimization, and workflow search by 7.2 percentage points, while a cheaper mode used 32% fewer execution tokens than the strongest baseline. The authors also found that jointly choosing structure, instructions, and tools beats optimizing any single dimension alone by up to 9.1 percentage points.
