Hugging Face releases harness adapters: safe RL inside pi, opencode and hermes
yb2698 · x · 2026-11-06
Hugging Face ships "harness adapters": one LoRA per model-harness pair, trained with RL on verifiable tasks inside unmodified agent harnesses (pi, opencode, hermes). Adapters make fewer tool calls while keeping accuracy; safeagents proxy blocks reward hacking by inspecting every tool call; training environments come from public coding sessions via Repo2RLEnvs; even the never-agent-trained TwIL-8B reasoning model learns.
More from coding & agent
- Claude Code v2.1.319 adds background MCP tasks, gateway spend tracking — claude-code-publisher[bot] · 2026-11-07
- 12 questions CEOs ask most about AI implementation, from 100+ exec calls — alex_verem · 2026-11-07
- Mold the Slop: carving a third road between embracing AI slop and micromanaging it — dscape · 2026-11-07
- Theo mocks startups chasing 6% token savings that make agents dumber — prateekj · 2026-11-07
- 10 mistakes CTOs make when implementing AI: demos tell you very little — alex_verem · 2026-11-07
- Click-to-summarize UI built on DocETL rewrite directives tames Anthropic's 421-page S-1 — sh_reya · 2026-11-07