WHALE (Joint Harness-Weight Optimization)

"A Simple Recipe for Joint Harness-Weight Optimization." KRAFTON AI, with Chelsea Finn, argues that an agent's performance depends jointly on its weights and on the executable harness that manages context and control flow, so optimizing either alone leaves the system bottlenecked by the frozen counterpart: weight updates change which harness works, harness updates change which capabilities are exposed. WHALE alternates two phases, online rejection-sampling fine-tuning of the model under the current harness, then a Meta-Harness search for a better harness under the updated model, with fixed or adaptive-patience switching. Experiments run Qwen3.5-2B and 4B agents on seven search-QA benchmarks, AIME 2024/25 with a Python tool, and Lichess puzzles. Code (verl vendored) is Apache 2.0; the paper posted August 31 with a blog on September 3.