Hugging Face Open-Sources Ultimate Guide to Multi-Harness RL
Hugging Face's post-training team (Lewis Tunstall et al.) released a fully open-source guide to multi-harness RL, motivated by a key observation: the same model with the same weights performs dramatically differently across agent harnesses (Claude Code, Codex, OpenCode, Mini-SWE-Agent). Several commentators (Saboo Shubham, Adithya S K) called it a must-read practical guide.
Confirmed
- Score gaps are large: untrained LFM2.5-2.6B solved 62% of tasks in Mini-SWE-Agent but only 33% in another harness, a 29-point gap.
- RL training inside the harness lifted LFM2.5 from 42% to 54% while reducing tool calls by 31%.
- The guide is fully open-source; Lewis notes every agent has a custom harness (e.g., Pi + extensions) and models often lose points when switching harnesses—one key trick is to train inside the environment without modifying the harness itself.
Why it matters
- Agent benchmark scores depend heavily on the harness, so single-framework scores don't reflect true model capability; cross-harness evaluation and training are becoming necessary.
- The guide demonstrates a viable path to optimize models for specific environments with in-harness RL, improving scores while cutting tool-call overhead—directly useful for open-source agent model developers.
2026-10-01 ~ 2026-10-03 · 6 related posts
Primary sources
- Hugging Face post-training team publishes the ultimate guide to multi-harness RL — lewtun ·
- RL inside the harnesses: lifting LFM2.5 from 42% to 54% across four agent harnesses — _lewtun ·
- Same model scores 62% vs 33% across agent harnesses; Hugging Face open-sources multi-harness RL guide — huggingface ·
- Multi-harness RL guide: LFM2.5 jumps 42% to 54% with 31% fewer tool calls — SergioPaniego · 2026-10-01
- [source] RL inside the harnesses: lifting LFM2.5 from 42% to 54% across four agent harnesses — _lewtun · 2026-10-02
- Hugging Face publishes the ultimate guide to multi-harness RL — Saboo_Shubham_ · 2026-10-02
- [source] Same model scores 62% vs 33% across agent harnesses; Hugging Face open-sources multi-harness RL guide — huggingface · 2026-10-02
- [source] Hugging Face post-training team publishes the ultimate guide to multi-harness RL — lewtun · 2026-10-03
1 near-duplicate retellings: Saboo_Shubham_