HF details harness adapters: one LoRA per agent harness, every tool call vetted by safeagents
yb2698 · x · 2026-11-06
Hugging Face's new blog post details the "harness adapters" method: since the same model behaves differently across agent harnesses (pi, opencode, hermes), they train a small LoRA adapter per harness on top of a base model, letting users swap the model behind their local agent for the adapter trained in that agent.
Key points:
- Method: RL on verifiable tasks inside the harness, one adapter per model×harness, all served from one model
- Results: in smoke training, adapters make fewer tool calls than the base model on held-out evals while keeping or improving accuracy
- New environments: Repo2RLEnvs turns public coding sessions on the Hub into verifiable tasks, mixed with SmolDataEnvs
- safeagents: a proxy between harness and model checks every tool call with rule-based detectors; reward-hacking attempts are blocked and score 0
- TwIL-8B: a reasoning model never trained as an agent learns to use harnesses too; final runs are streamed live
Earlier attempts on LFM2.5-24B-A2B showed encouraging signals.
More from coding & agent
- The great decoupling of search: three kinds of AI crawlers need different controls — lilyraynyc · 2026-11-07
- Dev shows multi-agent crew auto-creating a Black Friday discount code end-to-end for $0.10 — pramodk73 · 2026-11-06
- Developer says Codex wrecked his server after a reboot request, costing $10,000+ — burny_tech · 2026-11-06
- Exo: Berkeley's agent runtime that edits its own harness, cutting model cost 35-61% — dair_ai · 2026-11-06
- Symbolic belief revision plus geometric projections lifts RAG reranking to 0.481 nDCG@10 — letandrewcook · 2026-11-06
- Free open-source AI Engineering from Scratch course teaches math-first — HowDevelop · 2026-11-06