Infini-AI-Lab Releases Hermit, an Open-Supply Harness Adapter Framework for AI Brokers

Infini-AI-Lab has launched Hermit, an open-source harness adapter framework for AI brokers. Hermit is a framework from Infini-AI-Lab that trains an exterior ‘harness adapter’ whereas the agent’s mannequin and harness keep frozen. The first launch, v0.1.0, shipped on October 25, 2026. CMU postdoc Haizhong Zheng introduced it on X on October 27. For builders, the pitch is straightforward: adapt Codex or Claude Code brokers to new duties with out fine-tuning or forking the harness.

TL;DR

Footprint: A Python framework. No parameters of its personal. Apache-2.0, pip set up hermit-harness.
Runs on: The offline Fast Begin wants no GPU, mannequin account or dataset. Actual runs use any API backend or a neighborhood vLLM, SGLang or Ollama server.
Efficiency: Reported positive factors over no-adapter baselines on OfficeQA, CL-Bench and ServeLearnBench, all with GLM 5.3 Flash.
Greatest: MWM adapter hits 0.795 accuracy on CL-Bench database exploration, vs 0.580 for the in-context baseline.
Worst: The prompt-only adapter scores 0.470 on the identical venue, beneath that 0.580 baseline.
Backside line: Greatest: one coaching loop for five harness evolution strategies. Worst: most outcomes are single runs, self-reported.

What’s a harness adapter?

A harness adapter is a trainable exterior harness. It really works like a plugin with coaching assist in-built. A mannequin adapter, reminiscent of LoRA, adapts a mannequin. A harness adapter adapts an current agent as an alternative.

The adapter can maintain directions, expertise, reminiscence, instruments, executable code and agent coordination logic. It may possibly even be an entire repository. Its specification says which elements are learnable and the way updates get checked. The bottom mannequin and harness by no means change.

The crew offers 3 causes for this design. Many harnesses are closed-source. Modifying open ones can break rigorously tuned element interactions. And a discovered adapter can transfer between suitable fashions and harnesses.

How does Hermit practice an adapter?

Hermit makes use of an RL-style coaching loop. A Coach runs a batch of duties by means of the performing agent with the present adapter. A choose scores the outputs. An Evolver reads the trajectories and suggestions, then proposes a brand new adapter. The Coach checks it towards the template and commits it as the subsequent model.

Each adapter model lives in a git tree. You may examine, roll again or ship any model. Learner applications can run sandboxed, which requires Linux with bubblewrap.

5 evolution strategies ship right this moment: GEPA, SkillOpt, MetaWorld Mannequin (MWM), Harness² and Continuous Harness. ACE and ReHarness are listed as coming quickly. You too can write your personal Evolver.

<div class="pn on" data-i="0">
<p class="lead">A <b>harness adapter</b> is a trainable exterior harness, like a plugin with coaching in-built. Click on any half to see what it does.</p>
<div class="dg">
<div class="field frz"><h4>Your agent</h4>
<span class="chip" data-d="The bottom mannequin. Hermit by no means updates its weights.">Mannequin</span>
<span class="chip" data-d="The present harness (for instance Codex or Claude Code). Hermit by no means edits it, so closed-source harnesses work too.">Harness</span>
</div>
<div class="plug" aria-hidden="true"></div>
<div class="field advert"><h4>Harness adapter</h4>
<span class="chip" data-d="Instruction recordsdata reminiscent of ADAPTER.md or AGENTS.md that the agent reads as context.">Directions</span>
<span class="chip" data-d="Reusable expertise, together with executable helper scripts.">Expertise</span>
<span class="chip" data-d="Notes and saved information the agent carries throughout duties.">Reminiscence</span>
<span class="chip" data-d="Additional instruments, such because the wm software MWM exposes for its world mannequin.">Instruments</span>
<span class="chip" data-d="Executable code. An adapter will be a whole repository.">Code</span>
<span class="chip" data-d="Coordination logic for a number of brokers, as within the Lean adapter that hyperlinks 5 Codex brokers.">Coordination</span>
</div>
</div>
<div class="data" id="mhx-info">Each adapter model is saved in a git tree, so you’ll be able to examine, roll again or ship anybody.</div>
</div>

<div class="pn" data-i="1">
<p class="lead">Hermit runs an <b>RL-style loop</b>, however the factor being up to date is the adapter, not the weights.</p>
<div class="loop">
<div class="st"><div class="n">STEP 1</div><div class="t">Run duties</div></div>
<div class="st"><div class="n">STEP 2</div><div class="t">Choose scores</div></div>
<div class="st"><div class="n">STEP 3</div><div class="t">Evolver proposes</div></div>
<div class="st"><div class="n">STEP 4</div><div class="t">Examine vs template</div></div>
<div class="st"><div class="n">STEP 5</div><div class="t">Commit model</div></div>
</div>
<div class="data" id="mhx-step">Press Play to step by means of one coaching iteration.</div>
<div class="ctl">
<button class="btn" id="mhx-play"> Play</button>
<button class="btn gh" id="mhx-next">Subsequent step</button>
<button class="btn gh" id="mhx-reset">Reset</button>
<span class="ver">Adapter model <b id="mhx-v">V0</b></span>
</div>
<p class="observe">Constructed-in Evolvers: GEPA, SkillOpt, MetaWorld Mannequin (MWM), Harness², Continuous Harness. ACE and ReHarness are listed as coming quickly.</p>
</div>

<div class="pn" data-i="2">
<p class="lead">Self-reported outcomes, all with <b>GLM 5.3 Flash</b> because the performing mannequin. Decide a benchmark.</p>
<div class="seg">
<button class="tab on" data-b="0">OfficeQA</button>
<button class="tab" data-b="1">CL-Bench database</button>
<button class="tab" data-b="2">ServeLearnBench (H)</button>
</div>
<div class="bars" id="mhx-bars"></div>
<p class="observe" id="mhx-bnote"></p>
</div>

<div class="pn" data-i="3">
<p class="lead">Practice as soon as, then <b>deploy</b> the adapter into an agent workspace or export it as a local plugin.</p>
<div class="time period"><span class="p">$</span> pip set up hermit-harness
<span class="p">$</span> hermit practice –config run.yaml
<span class="p">$</span> hermit eval –config run.yaml –adapter ultimate
<span class="p">$</span> hermit deploy –adapter DIR –workspace WS –host codex
<span class="p">$</span> hermit convert –adapter DIR –output OUT –target claude-code
<span class="c"># different convert targets: codex, pi, hermes</span></div>
<div class="hosts">
<span class="chip" data-d="Deploy goal and plugin export goal.">Codex</span>
<span class="chip" data-d="Deploy goal and plugin export goal.">Claude Code</span>
<span class="chip" data-d="Deploy goal and plugin export goal.">Pi</span>
<span class="chip" data-d="Plugin export goal.">Hermes</span>
</div>
<div class="data" id="mhx-dinfo">Backends by reference: Fireworks, OpenAI, Anthropic, DeepSeek, vLLM, SGLang, Ollama. The offline Fast Begin wants no GPU, mannequin account or dataset.</div>
</div>

<div class="ft"><span>Supply: github.com/Infini-AI-Lab/hermit · self-reported</span><span class="mtp">© Marktechpost</span></div>
</div>
<script>
(perform(){
var R=doc.getElementById(‘mtp-hermit-x’);if(!R)return;
perform ping(){strive{var h=R.offsetHeight+40;mother or father.postMessage({mtpEmbedHeight:h,peak:h,frameHeight:h},’*’);}catch(e){}}
var tabs=R.querySelectorAll(‘.tabs .tab’),pns=R.querySelectorAll(‘.pn’);
tabs.forEach(perform(t){t.addEventListener(‘click on’,perform(){var p=+t.getAttribute(‘data-p’);tabs.forEach(perform(x){x.classList.take away(‘on’)});t.classList.add(‘on’);pns.forEach(perform(x){x.classList.toggle(‘on’,+x.getAttribute(‘data-i’)===p)});if(p===2)drawBars(cur);setTimeout(ping,60);});});
perform wire(sel,out){R.querySelectorAll(sel).forEach(perform(c){c.addEventListener(‘click on’,perform(){R.querySelectorAll(sel).forEach(perform(x){x.classList.take away(‘sel’)});c.classList.add(‘sel’);doc.getElementById(out).textContent=c.textContent+’: ‘+c.getAttribute(‘data-d’);ping();});});}
wire(‘.dg .chip’,’mhx-info’);wire(‘.hosts .chip’,’mhx-dinfo’);
var steps=R.querySelectorAll(‘.st’),si=-1,ver=0,timer=null;
var msg=[‘The Trainer runs a batch of tasks through the acting agent with the current adapter.’,’Your judge scores each output and writes feedback.’,’An Evolver reads trajectories and feedback, then proposes a new adapter.’,’The Trainer checks the proposal against the adapter template.’,’The checked update is committed as the next adapter version.’];
perform go(i){si=i;steps.forEach(perform(s,okay){s.classList.toggle(‘on’,okay===i)});doc.getElementById(‘mhx-step’).textContent=msg[i];if(i===4){ver++;doc.getElementById(‘mhx-v’).textContent=”V”+ver;}}
perform subsequent(){go((si+1)%5)}
doc.getElementById(‘mhx-next’).onclick=perform(){cease();subsequent()};
doc.getElementById(‘mhx-play’).onclick=perform(){if(timer){cease();return}this.textContent=”❚❚ Pause”;subsequent();timer=setInterval(subsequent,1400)};
perform cease(){clearInterval(timer);timer=null;doc.getElementById(‘mhx-play’).textContent=” Play”}
doc.getElementById(‘mhx-reset’).onclick=perform(){cease();si=-1;ver=0;steps.forEach(perform(s){s.classList.take away(‘on’)});doc.getElementById(‘mhx-v’).textContent=”V0″;doc.getElementById(‘mhx-step’).textContent=”Press Play to step by means of one coaching iteration.”};
var B=[
{max:1,fmt:3,rows:[[‘No adapter’,.541],[‘Initial skill’,.576],[‘GEPA’,.580],[‘SkillOpt’,.600],[‘Continual Harness’,.634,1],[‘Harness²’,.651,1]],n:’Accuracy on 172 take a look at questions. GEPA and SkillOpt are technique of 4 trainings, examined frozen. Continuous Harness and Harness² study whereas streaming the take a look at set as soon as. Customary error about 3.7 factors.’},
{max:1,fmt:3,rows:[[‘Stateless baseline’,.400],[‘Prompt-only’,.470],[‘ICL baseline’,.580],[‘MWM’,.795,1]],n:’Accuracy on a 40-question stream over an undocumented, drifting SQLite database. Imply of 5 stream orders.’},
{max:100,fmt:1,rows:[[‘Blind’,4.5],[‘Prompt-only’,18.8],[‘SkillOpt (leaderboard)’,22.0],[‘CH (leaderboard)’,28.7],[‘MWM’,45.3,1],[‘Oracle (policy given)’,89.2]],n:’retail_l3: 624 serving requests, 537 held-out checks. H = rating on checks that want the hidden coverage. One run every.’}];
var cur=0;
perform drawBars(okay){cur=okay;var d=B[k],el=doc.getElementById(‘mhx-bars’);el.innerHTML=”;d.rows.forEach(perform(r){var row=doc.createElement(‘div’);row.className=”row”+(r[2]?’ hl’:”);row.innerHTML='<span class="lb">’+r[0]+'</span><span class="tr"><span class="fl" fashion="show:block"></span></span><span class="v">’+r[1].toFixed(d.fmt)+'</span>’;el.appendChild(row);var f=row.querySelector(‘.fl’);setTimeout(perform(){f.fashion.width=(r[1]/d.max*100)+’%’},40);});doc.getElementById(‘mhx-bnote’).textContent=d.n;setTimeout(ping,60);}
R.querySelectorAll(‘.seg .tab’).forEach(perform(b){b.addEventListener(‘click on’,perform(){R.querySelectorAll(‘.seg .tab’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);drawBars(+b.getAttribute(‘data-b’));});});
drawBars(0);
window.addEventListener(‘load’,ping);window.addEventListener(‘resize’,ping);setTimeout(ping,300);
})();
</script>
</physique></html>”>

(perform(){var f=doc.getElementById(“mtp-hermit-frame”);window.addEventListener(“message”,perform(e)e.supply!==f.contentWindow)return;var h=e.information&&e.information.mtpEmbedHeight;if(h&&h>100)f.fashion.peak=h+”px”;);})();

What are the benchmark outcomes?

All runs beneath use GLM 5.3 Flash because the performing mannequin. The numbers come from the repo’s instance READMEs and are self-reported.

OfficeQA (recipe) asks 172 take a look at questions on U.S. Treasury Bulletins, scored by precise match:

Methodology	Accuracy
No adapter	0.541
Preliminary talent	0.576
GEPA	0.580
SkillOpt	0.600
Continuous Harness	0.634
Harness²	0.651

GEPA and SkillOpt practice on 50 questions and are examined frozen. Continuous Harness and Harness² as an alternative study whereas streaming the take a look at set as soon as. The crew studies a +7.6 level acquire for Harness² over its reference run, with a 95% interval of +1.7 to +13.4.

CL-Bench (database, poker) checks stateful environments from the Continuous Studying Bench paper. On database exploration, MWM reaches 0.795 accuracy. The benchmark’s in-context baseline scores 0.580. On exploitable poker, MWM wins 3.08 massive blinds per hand, vs 2.51 for in-context studying.

ServeLearnBench (recipe) streams 624 retail assist requests underneath a hidden, shifting coverage. MWM scores 45.3 on hidden-policy checks. The perfect studying technique on the leaderboard, CH, scores 28.7. An oracle given the coverage scores 89.2.

How does Hermit examine with different agent optimizers?

Function	Hermit	GEPA	ACE	Agent Lightning
What it trains	Exterior harness adapter	Any textual content parameter (prompts, code)	Context playbook	Prompts and mannequin weights
Adjustments mannequin weights	No	No	Not disclosed	Sure, by way of RL or SFT
Strategies	5 in-built, customized Evolvers	Reflective evolution, Pareto choice	Generator, Reflector, Curator	RL, immediate optimization, SFT
Deploy targets	Codex, Claude Code, Pi; plugin export	DSPy (`dspy.GEPA`)	Not disclosed	LangChain, OpenAI Brokers SDK, AutoGen, CrewAI and extra
License	Apache-2.0	MIT	Apache-2.0	MIT
Key consequence	0.795 on CL-Bench database (self-reported)	AIME 2025: 46.6% to 56.6% (GPT-4.1 Mini)	+10.6% on AppWorld	Not disclosed

How do you put in and deploy it?

Set up from PyPI with pip set up hermit-harness, or clone the repo for examples. The CLI covers verify, practice, eval, deploy and convert. Fashions are referenced as <backend>/<mannequin id>. Supported backends embrace Fireworks, OpenAI, Anthropic, DeepSeek, vLLM, SGLang and Ollama.

hermit deploy writes a workspace for Codex, Claude Code or Pi. hermit convert exports a educated adapter as a local plugin for Codex, Claude Code, Pi or Hermes. A completed Lean collaboration adapter coordinates 5 Codex brokers proving one theorem collectively.