Jason Weston: Harness Optimization Hits 56.0, Distillation Adds 5.5 More

jaseweston · x · 2026-11-17

Jason Weston shares two experimental results for improving agent performance:

1. Harness Optimization (discover): a meta agent reads trajectories, finds where memory use breaks, and rewrites LOAD/OFFLOAD; the best candidates (5 runs each) become parents. Average test score: 56.0 optimized vs 44.6 best hand-written vs 36.1 generic.

2. Harness Distillation (demonstrate + distill): run the model with its optimized prompts, swap them back to generic while keeping its responses, then LoRA fine-tune. At test time with the generic prompt: 43.1 vs 37.6 for SFT on its own generic runs.

The idea is to bake the behavior induced by a good harness into the weights, so a plain prompt at inference still scores higher.

Related event: Jason Weston Boosts Agent Memory via Harness Optimization and Distillation(2 posts)→

Original post →

More from coding & agent

coding & agent channel →