A CMU lab packed eleven other people's papers into one framework and got 460 stars in two weeks

Hermit is a Python framework that aims to let agents like Codex and Claude Code learn to revise their own instructions, skills and tools from task feedback. It didn't invent a new algorithm; it unifies eleven other people's methods into one training loop. Thanks to the lab's launch and a media repost, it got 463 stars in two weeks. The catch is that much of this attention came from endorsement and track heat, which you can't copy. What you can learn is the position it chose, the name it gave itself and its launch timing.

The Oct 26 thread

On Oct 26, 2026, the lab account @InfiniAILab at Carnegie Mellon University's Beidi Chen group posted a 5-tweet thread on X announcing the open-sourcing of Hermit, with the hashtags #HarnessEvolution and #RSI, and roughly 10.4k views. The same day, lead author Haizhong Zheng posted a long essay, "Open-Sourcing Hermit," on X and LinkedIn. Beidi Chen reposted it with a very shareable line: "An agent that makes the same mistake 100 times is just automating déjà vu."

That same day, Chinese AI blogger @dongxi_nlp also recommended it, with about 7k views. On GitHub, v0.1.0 was released at 10-26 in the early morning, and two days later the repo had 276 stars.

What it is

First, two terms. The layer around an agent (system prompt, AGENTS.md, skill files, tools, memory) is called the harness. "Harness evolution" means letting the agent run tasks, take feedback, and then automatically revise this layer, with the model itself left completely untouched.

Hermit borrows an analogy everyone understands: just as a model can take a LoRA-style adapter, a harness can take a "harness adapter." Neither the original model nor the harness is changed; an add-on layer sits beside them and is trained with a reinforcement-learning-like loop. Each trained version is saved in git so you can inspect and roll it back, and once trained it can be deployed directly as a Codex or Claude Code workspace or plugin.

It supports eleven methods, including GEPA, SkillOpt, ACE, Meta-Harness, Continual Harness and RRSI (README, 11-06), nearly all algorithms from other people's papers. Hermit's job is to provide a unified interface so researchers can compare these methods in the same codebase or write their own.

Why these two weeks

First, the people behind it carry weight. Beidi Chen's group is well known in inference acceleration (MagicDec, TriForce, Sequoia). The lab's official account, a lead-author essay, the advisor reposting with a punchy line, and a collaborator reposting again together ensured the launch was seen across the research community.

Second, it landed during a paper pile-up. From 2026 9–October, papers on "agents improving themselves" came one after another: RRSI, Google's self-improving harness, Salesforce's co-evolution of harness and model, Harness Continual Learning... Each paper came with its own code, and none could be compared with the others. The "unified framework" spot was empty, and Hermit stepped into it.

Third, it's easy to write about. "Model adapter → harness adapter" explains it in one sentence. The README has a GIF showing the add-on layer being trained from V0 to V5 with scores climbing, and quick start needs no API key and no GPU. On Oct 29, MarkTechPost wrote a write-up following that framing and shared it on X, and stars rose by 62 and 58 over the next two days, its fastest growth. After that, daily.dev, Medium, LinkedIn and several YouTube videos (one titled "AI AGENTS CAN NOW EVOLVE THEMSELVES… MEET HERMIT!") picked it up in turn.

(GitHub star count, recorded daily)

I couldn't find any discussion of it on HN or Reddit, and it never ran ads. The whole spread went: lab → research community → an AI media outlet → secondary reposts.

How many people are left once the excitement fades

Not much. 11-04 shipped v0.2.0 with a feature useful to ordinary devs: "run training on your Codex subscription, no API key needed." Star growth barely moved, and after that it added only single digits a day. The hermit-harness package on PyPI got 153 downloads in the last week and 5 in the last day (pypistats, 11-06). Against 463 stars, only a small share of people who starred it actually installed it (inference: many install from source, which keeps PyPI counts low).

Another signal makes the point clearer. rsi-gear/gear, with almost the same positioning (calling itself "the scikit-learn of RSI"), was created by August too. It gained 70 and 54 stars on 10-29 and 10-30 and now sits at 417 stars, close to Hermit. The two are independent, so neither copied the other, but they rose on the same days. That suggests the attention came mainly from the topic itself, not anything unique to Hermit (whether both peaks came from the same roundup is unverified).

Who else is doing the same thing

The single-algorithm slot already has projects with tens of thousands of stars. The "unified framework" slot has only Hermit and gear, both at a few hundred stars, with no clear winner yet. Codex and Claude Code themselves only offer interfaces like AGENTS.md, skills, and plugins, with no training loop built in. That's why Hermit can plug into them.

Three things worth copying

Pick the position: build the aggregation layer, don't build another algorithm. When papers in one area pile up, each with its own code and no way to compare, the first project that puts them all behind one interface becomes the place people have to go to compare methods.

Pick the name: borrow a word people already know. "Harness adapter" borrows the intuition of LoRA / adapters. Readers don't have to learn a new concept, and media can write it up straight from the README.

Launch rhythm: costs almost nothing. An official thread, a long post from the author (one each for X and LinkedIn), a memorable one-liner, a repost by a Chinese blogger, then wait two or three days for AI media to pick it up. The README also links a WeChat group and Discord, so the Chinese community can join too.

If it were you

Verdict: Hermit took off partly because of the CMU lab's name and the advisor's reposts, and partly because it landed in the weeks when papers were piling up. You can't borrow either. Its results are modest anyway: 460 stars, 150 downloads a week, and gear in the same spot is about the same. Building another "unified harness evolution framework" on the same idea would already be the third.

Who it's for: developers building agent tools who can read papers and are willing to carefully reproduce other people's methods. What's worth borrowing is the thinking. Apply it to a different direction that is just starting to pile up and has no unified tool yet, and build "a unified framework for that class of papers."

Who it's not for: people with no account in the scene and no research background. Stars on projects like this depend on trust within the research community. However good the README, nothing takes off without people reposting it.

How to do it: two paths. One is to fill the part Hermit didn't do well. What it's missing is the trained adapters themselves: nobody has publicly shared a finished, trained result on a real task. Build a repo that uses Hermit or gear to train a few adapters for Codex / Claude Code on specific scenarios people actually have (for example, migrating code in a particular framework, or fixing a certain class of tests), with before-and-after score comparisons, and make them directly installable. That's easier to earn stars than writing another framework (inference). The other is to copy the playbook in a new direction: find the next sub-area where papers are piling up, be the first to put three to five papers' methods behind one interface, and pick a name that borrows an existing concept.

How to get users: start with AI researchers on X and Chinese AI bloggers. Write the README so media can copy it directly (one analogy, one animated GIF, one command that runs without a key), and actively send it to outlets like MarkTechPost. HN and Reddit have reacted little to projects like this. Hermit has zero discussion threads there.

How early you need to be: the heat of this harness-evolution wave peaked at the end of October, and post-launch star growth has already dropped to single digits a day. If you want to pursue the "finished adapter" route, you should move before Hermit or gear releases an official adapter collection itself. That gives you roughly one to two months (estimate).

Signs you got it wrong: (1) Hermit gets a paper or a conference talk, which brings another wave of several hundred stars. That would mean endorsements matter more than estimated here, and following up gets harder. (2) After one to two months, PyPI weekly downloads are still over a hundred, yet no one in the issues shares an adapter they trained. That would mean the direction is only being played with by researchers, and finished adapters wouldn't be used either. (3) Codex or Claude Code adds built-in "auto-update AGENTS.md / skills from feedback," and then an external framework layer like this isn't really needed.