Infini-AI-Lab released Hermit v0.1.0, an open-source Apache-2.0 Python framework that trains external harness adapters for AI agents while keeping the base model and harness frozen. Installed via pip install hermit-harness, it supports API backends (Fireworks, OpenAI, Anthropic, DeepSeek) and local servers (vLLM, SGLang, Ollama), and deploys adapters as native plugins for Codex, Claude Code, and Pi. Hermit uses an RL-style training loop: a Trainer executes tasks with the current adapter, a Judge scores outputs, and an Evolver proposes the next adapter version, all tracked in a git tree. Five evolution methods ship—GEPA, SkillOpt, MetaWorld Model (MWM), Harness², and Continual Harness—with ACE and ReHarness planned. Benchmarks use GLM 5.3 Flash and are self-reported, mostly single runs. On OfficeQA (172 U.S. Treasury questions), Harness² reaches 0.651 accuracy versus 0.541 no-adapter baseline (+7.6 points, 95% CI +1.7 to +13.4). On CL-Bench database exploration, MWM achieves 0.795 accuracy against a 0.580 in-context baseline; on exploitable poker, MWM wins 3.08 big blinds/hand versus 2.51. On ServeLearnBench (624 streaming retail requests), MWM scores 45.3 on hidden-policy tests versus the leaderboard best CH at 28.7 (oracle 89.2). Prompt-only adapter underperforms at 0.470 on CL-Bench. The CLI covers check, train, eval, deploy, and convert; a Lean collaboration example coordinates five Codex agents to prove a theorem.