Kangwook Lee: agent harness improvement can happen inside rollouts once agents are capable enough
Kangwook_Lee · x · 2026-09-04
Responding to a question about his paper, Kangwook Lee explains that the distinction between a "meta-harness" (the one that trains) and "harness" (the one being trained) was a deliberate setup choice, not a necessity.
Once that distinction goes away and agents become capable enough, harness improvement can happen as part of the rollouts themselves — in-context harness adaptation.
Related event: WHALE Paper Proposes Alternating Weight and Harness Optimization(7 posts)→
More from Research
- OpenAI researcher: molecular discovery must jointly optimize performance and manufacturability — shyamalanadkat · 2026-10-21
- First Passage Attention: path extrema enable perfect long-range retrieval with 25% cache — Yuxuan Liu · 2026-10-21
- Flattening weakest neuron activations barely hurts loss but devastates memorized text recall — ParshinShojaee · 2026-10-21
- Holon-S1 Predicts Genes and Biomarkers Directly From Pathology Slides Across 16 Organ Types — iScienceLuvr · 2026-10-21
- OpenVLHarness turns any MLLM into a visual reasoner, +9.9 points across 22 datasets — yining_hong · 2026-10-21
- Pedagogical RL from MIT and Laude Institute teaches models to generate learnable reasoning trajectories — lateinteraction · 2026-10-21