Multi-harness RL guide: LFM2.5 jumps 42% to 54% with 31% fewer tool calls
SergioPaniego · x · 2026-10-10
adithyask's team released 'The Ultimate Guide to Multi-Harness RL' — 50K+ reads in a week, 3 days trending on Twitter, #2 on Hugging Face. Key insight: the same model behaves differently across agent harnesses like Claude Code, Codex, and OpenCode. Their open method trains any model with RL on any task set inside the harnesses people actually use, without changing a single line of harness or training code. Trained across four harnesses, LFM2.5-2.6B went from 42% to 54% with 31% fewer tool calls.
More from coding & agent
- Why Bun AOT compiles from .ts source: d.ts type declarations aren't enough — MarcJSchmidt · 2026-10-10
- WareTwin: Open-Source Real-Time 3D Digital Twin Simulating 20 Warehouse AMRs — rsasaki0109 · 2026-10-10
- Google adds "Code Comprehension" interview round: debug a real codebase with Gemini, and catch when it's wrong — burny_tech · 2026-10-10
- Our AI bill hit $11,400 with no attribution: a cautionary tale of unbounded retries and prompt bloat — vigilAPI · 2026-10-10
- Doubao canvas upgrade: one image grows into a 24,000px VI manual, e-commerce pages and decks — xiaohu · 2026-10-10
- Dev uses Codex to migrate a 5.5TB S3 setup to AWS — CtrlAltDwayne · 2026-10-10