WHALE co-optimizes model weights and the agent harness

WHALE (Weight-Harness Alternating LEarning, arXiv 2609.00196, Wed 2 Sep digest) alternates two phases: updating the model under the current harness (online rejection-sampling fine-tuning), then searching for a better harness under the updated model. The premise is that agent performance depends jointly on weights and executable harness code, so optimizing either in isolation leaves the system bottlenecked by the frozen counterpart — weight updates change which harness is effective, and harness updates change which model capabilities get exposed. Prior joint-adaptation work optimized weights and textual prompts but left the broader harness fixed.

This extends the harness-optimization line the KB has tracked since AutoSaddler (8/26) and HarnessLens (8/29) from prompt-and-tool search into weight territory. If you fine-tune agents for production, the practical warning is direct: a harness tuned for the old checkpoint can silently waste the gains of a new one, so harness re-tuning belongs in the same release cadence as model updates. The same digest also carries HarnessDev (2609.01437, LLMs creating and evolving their own harness) and HarnessEvolve (2609.00829, reference-trajectory learning for reliable self-evolution) — four harness-focused papers in three digests.