On 2026-11-06 Hugging Face Harness Hub released Harness Adapters: one rank-32 LoRA per base model and coding harness (pi, OpenCode, Hermes), trained with GRPO on verifiable tasks inside unmodified harnesses. Vendor smoke evals claim fewer tool calls with equal or better accuracy. A safeagents proxy rule-checks every tool call and blocks reward hacks. Bases include LiquidAI LFM2.5-24B-A2B, LFM3-1.7B, and experimental TwIL-8B.

Key Takeaways

  • ✓Mechanics: one LoRA per model x harness (rank 32, all-linear), GRPO-trained inside unmodified pi / OpenCode / Hermes
  • ✓Safety: safeagents proxy rule-checks tool calls (undeclared tools, test edits, reward-file writes, answer-source fetches); blocked rollouts score 0
  • ✓Vendor smoke: fewer tool calls than the base with equal or better accuracy; one vLLM serves named adapters
  • ✓Envs: SmolDataEnvs plus 230 Repo2RLEnvs tasks (repair/build/extend) rebuilt from public Hub trajectories
  • ✓Open: blog, safeagents, training recipe, and live dashboard; follow runs at https://harness-hub-harness-adapters-rl.hf.space/
Hugging Face Ships Harness Adapters: RL-Trained LoRAs Inside Unmodified pi / OpenCode / Hermes, with safeagents Blocking Reward Hacks
🖼️Official Media
Click to view high-res
⚡

Key Decision Metrics at a Glance

TL;DR Verdict
🏆 Benchmark RankRank
Safety: safeagents proxy rule-checks tool calls (undeclared tools, test edits, reward-file writes, answer-source fetches); blocked rollouts score 0
Full Leaderboard→
💰 Cost & PlansCompare
API & Subscription
Compare Plans↗
🛠️ Ecosystem & IDERules
Hugging Face · Harness Adapters · LoRA
Setup IDE Rules↗
🧭

Turn your technical choice into a development budget

Compare 40 dev plans & simulate token costs vs $20/mo subscriptions

⚔️ Compare Model→💰 Compare plans & limits↗
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
📝
Hugging Face blog: Harness adapters
huggingface.co/blog/harness-hub/harness-adapters
↗
🐙
huggingface/safeagents
github.com/huggingface/safeagents
↗
🚀
Live training dashboard (HF Space)
harness-hub-harness-adapters-rl.hf.space/
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points

The same base weights behave differently across coding harnesses. Full fine-tunes are costly, and RL invites reward hacks. Hugging Face Harness Hub trains a small LoRA per harness and adds a safety gate.

Architecture Highlights & Internals

One rank-32 all-linear LoRA per base x harness, GRPO-trained inside unmodified pi / OpenCode / Hermes. One vLLM serves named adapters. safeagents reverse-proxies every tool call and blocks undeclared tools, test edits, reward-file writes, and answer-source fetches; blocked rollouts score 0.

Authoritative Benchmarks & Measured Scores

No full independent leaderboard. Blog smoke / held-out: fewer tool calls with equal or better accuracy. Repo2RLEnvs skew to build (~253 of ~270); SmolDataEnvs mixed in. TwIL-8B smoke blocked at step 0 for probing an HF bucket. Vendor/author-stated.

Developer Hands-on Guide

Attach the adapter for your harness to your local base without changing the harness. DIY: save trajectories, scrub/publish via safeagents, rebuild with Repo2RLEnv, train with GRPO. See external_links.

Action HubReady to adopt this in production?

Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.

⚔️ Launch VS Compare→💳 View Plans↗
Up NextZhejiang University & Tencent Open-Source KnowGPO for Deep Search Agents: Knowledge-Grouped RL Outperforms 32B Models, Scoring 40.2 on GAIA
Read Now→