RL-trained harness adapters drop for LFM3, LFM2.5 and TwiL-8B to boost agentic coding
yb2698 · x · 2026-11-09
harness-hub released new harness adapter collections on Hugging Face: LFM3-1.7B, LFM2.5-24B (look for the -v2 suffix) and TwiL-8B, all trained with RL on adapters for more efficient agentic development, with pi, hermes and opencode variants plus GGUF quants.
The author also shared findings from testing: many of the 179 blocked tool calls were false positives, such as harmless scripts run against benign hosts. Two fixes are planned: loosening safeagents plugin conditions for those hosts, and routing blocked calls to top models as a decision layer.
More from coding & agent
- CS spent 70 years killing nondeterminism; in 2026 we set temperature=0.8 — korymath · 2026-11-09
- Open-source Bello claims 3x more usage from your coding agent at same quality — nikola_mr64990 · 2026-11-09
- Nominal raises $240M Series C at $3B valuation for AI hardware engineering — graceisford · 2026-11-09
- HF's Niels Rogge releases 'Getting Started with Local AI' slides: quantization, MoE, local agents — NielsRogge · 2026-11-09
- A 24/7 pixel-art TV channel for an invented sport, built entirely by Claude Code — spite_as_an_arrow · 2026-11-09
- GPT-4o silently drops schema invariants at turn 18 in 50-turn agent stress test, token burn hits 24.2x — ParrotIntegrated · 2026-11-09