The Infini-AI-Lab has released Hermit, an open source harness adaptation framework for artificial intelligence agents. Its first version, v0.1.0, was released on October 25, 2026, and researcher Haizhong Zheng, a postdoctoral fellow at CMU, announced it on X on October 27. The stated goal is aimed at developers: to adapt Codex or Claude Code agents to new tasks without retraining the model or modifying the harness.
- Infini-AI-Lab releases Hermit, an open source framework that adapts Codex and Claude Code agents to new tasks without retraining the model.
- Hermit trains an external adapter that adds instructions, skills, memory, or tools, while the model and harness stay fixed.
- A trainer, a judge, and an evolver run a reinforcement-learning-style loop that improves the adapter version after version.
- The reported gains come from the authors and mostly from single runs, so they need independent confirmation.
An external adapter rather than a modified model
Hermit trains a harness adapter, that is, an external and trainable component that attaches to an existing agent. The harness here refers to the set of instructions, tools, and rules that govern an agent’s operation. Both the base model and the harness remain fixed. The adapter can contain instructions, reusable skills, memory, tools, executable code, or even the coordination logic of multiple agents, and it can even take the form of a complete repository.
According to the team, this choice is based on three reasons. Many harnesses are closed, modifications to an open harness can break carefully tuned interactions, and a trained adapter can be transferred between compatible models and harnesses. Unlike a model adapter like LoRA, which modifies the model itself, this one acts on the agent.
A five-step training loop
Hermit relies on a reinforcement learning type training loop. A Trainer has the agent execute a batch of tasks with the current adapter, then a judge scores the results. An Evolver reads the trajectories and feedback and proposes a new adapter. The Trainer verifies this proposal against a template before saving it as the next version.
Each version is stored in a git tree, allowing inspection, rollback, or distribution of any version. Learning programs can be run in a sandbox, which requires Linux with bubblewrap. Five evolution methods are available: GEPA, SkillOpt, MetaWorld Model (MWM), Harness², and Continual Harness. ACE and ReHarness are announced as coming soon, and users can write their own Evolver.
Self-reported results
All tests presented use GLM 5.3 Flash as the acting model. The figures come from the README files of the repository examples and are self-reported. On OfficeQA, which poses 172 questions on US Treasury Bulletins with an exact match criterion, accuracy rises from 0.541 without an adapter to 0.651 for Harness². On CL-Bench, MWM reaches 0.795 in database exploration, compared to 0.580 for the in-context learning baseline. On ServeLearnBench, which handles 624 support requests in a hidden policy setting, MWM scores 45.3, versus 28.7 for CH and 89.2 for an oracle that knows the policy. Only a single run was performed for this last test, and most published results also come from a single run.
On the deployment side, Hermit installs with pip install hermit-harness. The command hermit deploy writes a workspace for Codex, Claude Code, or Pi, while hermit convert exports a trained adapter as a native plugin for Codex, Claude Code, Pi, or Hermes. Models are designated by the form /, including Fireworks, OpenAI, Anthropic, DeepSeek, vLLM, SGLang, and Ollama. The repository also features a collaboration adapter for Lean, which coordinates five Codex agents jointly tasked with proving the same theorem.
