Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI
  • AWS Routes HyperPod Inference Throughout Areas From One Endpoint – Unite.AI
  • DailyAI Briefing: A Lean-checked Riemann proof, plus 4 extra
  • Anthropic Shares Drop 8.5% in Second Session Throughout AI Selloff – Unite.AI
  • OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min
  • Copywriting: The Full Information | Copyblogger
  • AI Governance in 2027: Accountability Strikes Into the Workflow – Unite.AI
AI-trends.todayAI-trends.today
Home»Tech»Google Open-Sources Harness2, a Recursive Agent Harnessing Framework

Google Open-Sources Harness2, a Recursive Agent Harnessing Framework

Tech By Gavin Wallace18/10/20269 Mins Read
Facebook Twitter LinkedIn Email
A Coding Implementation to Build an AI Agent with Live
A Coding Implementation to Build an AI Agent with Live
Share
Facebook Twitter LinkedIn Email

Google Cloud AI Research has open-sourced Harness². It’s a recursive agent harnessing framework that improves the code and recordsdata round a frozen mannequin. For builders, the pitch is easy. Your agent might get higher with no higher mannequin.

TL;DR

  • Measurement: A framework, not a mannequin. 0 parameters skilled; 8 editable harness parts. Model 0.1.0.
  • Runs on: Linux with Python 3.12 or 3.13 and uv. Mannequin runs name fashions by means of Vertex AI. GPU requirement: Not disclosed.
  • Efficiency: 1 refinement step improved all 4 mannequin and harness setups on all 3 benchmarks.
  • Greatest: GLM-5 underneath OpenCode rose from 64.2 to 75.7 on WorkBuddy-Bench (+11.5 factors).
  • Worst: GLM-5 underneath Codex gained just one.2 factors on JobBench (23.6 to 24.8).

What’s Harness²?

Harness² is a plug-and-play framework from Google Analysis that refines an agent’s harness throughout deployment. A harness is every little thing across the mannequin: prompts, tools, memory, skills and orchestration. Harnessing is the method of optimizing that system. Mannequin weights by no means change.

The goal setting is the open world. Duties arrive unseen, and no verifier says whether or not a solution is correct. The research team argues that almost all self-improvement strategies assume labels or take a look at circumstances. Harness² is constructed to work with out them.

What are the 8 harness parts?

Harness² exposes 8 editable components, every mounted as a file:

  1. System immediate
  2. Guardrails
  3. Reminiscence
  4. Device steering
  5. Subagents
  6. Abilities
  7. Scripts
  8. Plugins (middleware)

They cowl context building, the device interface, delegation and habits contained in the device loop. The format extends the 7-file break up from earlier Agentic Harness Engineering work by 1 part.

How does recursive agent harnessing work?

The framework runs two linked ranges of recursion:

  • Job-level recursion makes use of a contrastive Suggest, Probe, Compose step. First, the agent runs the duty as soon as with the bottom harness. An editor then proposes 2 candidates. The invariant candidate targets basic agent failures reminiscent of planning or device use. The area candidate targets information and procedures for the present area. Each candidates run on the identical process. The composer compares their trajectories in opposition to the reference and retains solely edits backed by noticed habits.
  • Harnessing-level recursion occurs after every process. A reflector shops the adopted harness plus an EXPERIENCE.md file of classes. This library guides future proposals. Every new process nonetheless begins from the bottom harness.

The finances ok scales the search. Parallel mode probes ok invariant and ok area candidates, then makes ok compositions. Sequential mode chains ok refinement steps.

<iframe id="mtp-h2x-frame" title="Harness2 interactive explainer" loading="lazy" type="width:100%;height:600px;border:0;display:block;background:#0f1115;border-radius:16px;" scrolling="no" srcdoc="<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Harness2 Explainer</title><style>html,body{margin:0;background:#0f1115}body{padding:0}</style></head><body><div id="mtp-h2x">
<type>
#mtp-h2x{–bg:#0f1115;–card:#171a21;–line:#2a2f3a;–tx:#e8eaed;–mut:#9aa0a6;–b:#4285F4;–r:#EA4335;–y:#FBBC04;–g:#34A853;–mtp:#76B900;
max-width:880px;margin:0 auto;background:var(–bg)!vital;coloration:var(–tx)!vital;border:1px stable var(–line)!vital;border-radius:16px;padding:22px 22px 16px;font-family:Inter,"Google Sans",Roboto,system-ui,-apple-system,sans-serif;box-sizing:border-box;line-height:1.5}
#mtp-h2x *{box-sizing:border-box}
#mtp-h2x hr,#mtp-h2x p:empty,#mtp-h2x del,#mtp-h2x s{show:none!vital}
#mtp-h2x .hd{show:flex;align-items:middle;hole:10px;flex-wrap:wrap;margin-bottom:4px}
#mtp-h2x .dots{show:flex;hole:5px}
#mtp-h2x .dots i{width:10px;peak:10px;border-radius:50%;show:block;animation:h2pulse 2.4s infinite}
#mtp-h2x .dots i:nth-child(2){animation-delay:.3s}#mtp-h2x .dots i:nth-child(3){animation-delay:.6s}#mtp-h2x .dots i:nth-child(4){animation-delay:.9s}
@keyframes h2pulse{0%,100%{remodel:scale(1);opacity:1}50%{remodel:scale(.6);opacity:.5}}
#mtp-h2x h3{margin:0!vital;font-size:21px;font-weight:700;coloration:var(–tx)!vital;letter-spacing:-.01em}
#mtp-h2x h3 sup{coloration:var(–b)}
#mtp-h2x .sub{coloration:var(–mut)!vital;font-size:13.5px;margin:2px 0 14px}
#mtp-h2x .tabs{show:flex;hole:6px;flex-wrap:wrap;border-bottom:1px stable var(–line);margin-bottom:16px}
#mtp-h2x .tab{background:clear!vital;border:0!vital;coloration:var(–mut)!vital;font:600 13.5px/1 inherit;font-family:inherit;padding:10px 12px;cursor:pointer;border-bottom:2px stable clear!vital;margin-bottom:-1px}
#mtp-h2x .tab.on{coloration:var(–tx)!vital;border-bottom-color:var(–b)!vital}
#mtp-h2x .pn{show:none;animation:h2in .35s ease}
#mtp-h2x .pn.on{show:block}
@keyframes h2in{from{opacity:0;remodel:translateY(6px)}to{opacity:1;remodel:none}}
/* pipeline */
#mtp-h2x .move{show:grid;grid-template-columns:repeat(4,1fr);hole:10px;place:relative}
#mtp-h2x .nd{background:var(–card)!vital;border:1px stable var(–line)!vital;border-radius:12px;padding:11px 12px;transition:all .35s;opacity:.38;min-height:92px}
#mtp-h2x .nd b{show:block;font-size:13.5px;coloration:var(–tx)!vital}
#mtp-h2x .nd small{show:block;coloration:var(–mut)!vital;font-size:12px;margin-top:3px}
#mtp-h2x .nd .tg{show:inline-block;font-size:10.5px;font-weight:700;letter-spacing:.06em;padding:2px 7px;border-radius:99px;margin-bottom:6px;coloration:#0f1115!vital}
#mtp-h2x .nd.lit{opacity:1;remodel:translateY(-2px);box-shadow:0 0 0 1px var(–c),0 8px 22px -10px var(–c)}
#mtp-h2x .lvl{font-size:11px;font-weight:700;letter-spacing:.08em;coloration:var(–mut)!vital;margin:12px 0 6px;text-transform:uppercase}
#mtp-h2x .ctl{show:flex;hole:8px;align-items:middle;margin-top:14px;flex-wrap:wrap}
#mtp-h2x .btn{background:var(–card)!vital;coloration:var(–tx)!vital;border:1px stable var(–line)!vital;border-radius:99px;padding:8px 14px;font:600 13px inherit;font-family:inherit;cursor:pointer}
#mtp-h2x .btn.pri{background:var(–b)!vital;border-color:var(–b)!vital;coloration:#fff!vital}
#mtp-h2x .be aware{background:var(–card)!vital;border:1px stable var(–line)!vital;border-left:3px stable var(–b)!vital;border-radius:10px;padding:11px 13px;margin-top:12px;font-size:13.5px;coloration:var(–tx)!vital;min-height:64px}
#mtp-h2x .prog{flex:1;min-width:120px;peak:4px;background:var(–line);border-radius:4px;overflow:hidden}
#mtp-h2x .prog i{show:block;peak:100%;width:0;background:linear-gradient(90deg,var(–b),var(–r),var(–y),var(–g));transition:width .4s}
/* parts */
#mtp-h2x .cg{show:grid;grid-template-columns:repeat(4,1fr);hole:8px}
#mtp-h2x .cp{background:var(–card)!vital;border:1px stable var(–line)!vital;border-radius:11px;padding:11px;cursor:pointer;transition:all .25s;text-align:left;coloration:var(–tx)!vital;font-family:inherit}
#mtp-h2x .cp:hover,#mtp-h2x .cp.on{border-color:var(–c)!vital;remodel:translateY(-2px)}
#mtp-h2x .cp em{font-style:regular;font-size:11px;font-weight:700;coloration:var(–c)!vital}
#mtp-h2x .cp b{show:block;font-size:13.5px;margin-top:2px}
#mtp-h2x .lg{show:flex;hole:14px;flex-wrap:wrap;font-size:12px;coloration:var(–mut)!vital;margin:10px 0 0}
#mtp-h2x .lg span:earlier than{content material:"";show:inline-block;width:9px;peak:9px;border-radius:3px;margin-right:5px;background:var(–c)}
/* bars */
#mtp-h2x .sel{show:flex;hole:6px;flex-wrap:wrap;margin-bottom:12px}
#mtp-h2x .chip{background:var(–card)!vital;border:1px stable var(–line)!vital;coloration:var(–mut)!vital;border-radius:99px;padding:6px 11px;font:600 12.5px inherit;font-family:inherit;cursor:pointer}
#mtp-h2x .chip.on{coloration:#fff!vital;background:var(–b)!vital;border-color:var(–b)!vital}
#mtp-h2x .bm{show:grid;grid-template-columns:repeat(3,1fr);hole:14px}
#mtp-h2x .bx{background:var(–card)!vital;border:1px stable var(–line)!vital;border-radius:12px;padding:12px}
#mtp-h2x .bx h4{margin:0 0 10px!vital;font-size:13px;coloration:var(–tx)!vital}
#mtp-h2x .row{show:grid;grid-template-columns:62px 1fr 42px;align-items:middle;hole:8px;margin:7px 0;font-size:12px;coloration:var(–mut)!vital}
#mtp-h2x .tr{peak:12px;background:#222733;border-radius:6px;overflow:hidden}
#mtp-h2x .tr i{show:block;peak:100%;width:0;border-radius:6px;transition:width .9s cubic-bezier(.2,.8,.2,1)}
#mtp-h2x .row b{coloration:var(–tx)!vital;text-align:proper;font-size:12.5px}
#mtp-h2x .dl{font-size:12px;font-weight:700;coloration:var(–g)!vital;margin-top:4px}
#mtp-h2x .rk .row{grid-template-columns:120px 1fr 46px}
#mtp-h2x .fn{font-size:11.5px;coloration:var(–mut)!vital;margin-top:10px}
#mtp-h2x .ft{show:flex;justify-content:space-between;hole:8px;flex-wrap:wrap;border-top:1px stable var(–line);margin-top:16px;padding-top:10px;font-size:11.5px;coloration:var(–mut)!vital}
#mtp-h2x .ft a{coloration:var(–mtp)!vital;text-decoration:none;font-weight:700}
#mtp-h2x code{background:#222733!vital;coloration:var(–y)!vital;border:0!vital;padding:1px 5px;border-radius:4px;font-size:12px}
@media(max-width:640px){
#mtp-h2x{padding:16px 14px 12px}
#mtp-h2x .move,#mtp-h2x .cg{grid-template-columns:repeat(2,1fr)}
#mtp-h2x .bm{grid-template-columns:1fr}
#mtp-h2x h3{font-size:18px}
#mtp-h2x .rk .row{grid-template-columns:96px 1fr 42px}
}
@media(prefers-reduced-motion:scale back){#mtp-h2x *{animation:none!vital;transition:none!vital}}
</type>

<div class="hd"><div class="dots"><i type="background:#4285F4"></i><i type="background:#EA4335"></i><i type="background:#FBBC04"></i><i type="background:#34A853"></i></div><h3>Harness<sup>2</sup> defined</h3></div>
<div class="sub">Google's recursive agent harnessing: the harness modifications, the mannequin weights keep frozen.</div>

<div class="tabs">
<button class="tab on" data-t="p1">The way it works</button>
<button class="tab" data-t="p2">8 parts</button>
<button class="tab" data-t="p3">Benchmarks</button>
<button class="tab" data-t="p4">vs baselines</button>
</div>

<div class="pn on" id="h2-p1">
<div class="lvl">01 · Job-level recursion (inside 1 process)</div>
<div class="flow">
<div class="nd" type="–c:#9aa0a6"><span class="tg" type="background:#9aa0a6">REFERENCE</span><b>Run base harness</b><small>H0 executes the duty as soon as</small></div>
<div class="nd" type="–c:#4285F4"><span class="tg" type="background:#4285F4">PROPOSE</span><b>2 candidates</b><small>Invariant + domain-specific edits</small></div>
<div class="nd" type="–c:#EA4335"><span class="tg" type="background:#EA4335">PROBE</span><b>Run each</b><small>Identical process, similar mannequin, examine habits</small></div>
<div class="nd" type="–c:#FBBC04"><span class="tg" type="background:#FBBC04">COMPOSE</span><b>Merge what labored</b><small>Preserve edits backed by proof</small></div>
</div>
<div class="lvl">02 · Harnessing-level recursion (throughout duties)</div>
<div class="flow">
<div class="nd" type="–c:#4285F4"><span class="tg" type="background:#4285F4">EXECUTE</span><b>Last harness</b><small>Deliverable returned</small></div>
<div class="nd" type="–c:#34A853"><span class="tg" type="background:#34A853">REFLECT</span><b>Assessment the file</b><small>Which edits helped, and why</small></div>
<div class="nd" type="–c:#34A853"><span class="tg" type="background:#34A853">UPDATE</span><b>Enchancment library</b><small>Harness + EXPERIENCE.md</small></div>
<div class="nd" type="–c:#9aa0a6"><span class="tg" type="background:#9aa0a6">NEXT TASK</span><b>Restart at H0</b><small>Library guides new proposals</small></div>
</div>
<div class="note" id="h2-note">Press <b>Play</b> or <b>Step</b> to stroll by means of 1 Harness<sup>2</sup> enchancment step.</div>
<div class="ctl"><button class="btn pri" id="h2-play"> Play</button><button class="btn" id="h2-step">Step →</button><button class="btn" id="h2-reset">↺ Reset</button><div class="prog"><i id="h2-prog"></i></div></div>
<div class="fn">At ok=1: 1 reference + 2 probes + 1 remaining run per process. Parallel mode probes ok invariant and ok area candidates; sequential mode chains ok steps.</div>
</div>

<div class="pn" id="h2-p2">
<div class="cg" id="h2-cg"></div>
<div class="note" id="h2-cnote" type="border-left-color:#4285F4">Click on a part to see its runtime position.</div>
<div class="lg"><span type="–c:#4285F4">Context building</span><span type="–c:#EA4335">Device interface</span><span type="–c:#FBBC04">Orchestration</span><span type="–c:#34A853">Middleware</span></div>
</div>

<div class="pn" id="h2-p3">
<div class="sel" id="h2-sel"></div>
<div class="bm" id="h2-bm"></div>
<div class="fn">Single enchancment step (ok=1). Means over 3 runs. LAB = Partial cross fee (%). Supply: Harness<sup>2</sup> paper, Tables 1 and a couple of.</div>
</div>

<div class="pn" id="h2-p4">
<div class="sel" id="h2-sel2"></div>
<div class="bx rk" id="h2-rk"></div>
<div class="fn">Gemini 3.5 Flash executor. Bar axis begins at 50. Meta-Harness is optimized offline on separate coaching duties. Scores as measured within the Harness<sup>2</sup> paper.</div>
</div>

<div class="ft"><span>Sources: arXiv 2610.14224 · github.com/google-research/harness2</span><a href="https://www.marktechpost.com" goal="_blank" rel="noopener">© Marktechpost</a></div>

<script>
(perform(){
var R=doc.getElementById(‘mtp-h2x’);
perform $(s){return R.querySelector(s)}perform $$(s){return R.querySelectorAll(s)}
perform sz(){strive{guardian.postMessage({mtpH2xHeight:R.offsetHeight+40},’*’)}catch(e){}}
$$(‘.tab’).forEach(perform(t){t.onclick=perform(){$$(‘.tab’).forEach(perform(x){x.classList.take away(‘on’)});$$(‘.pn’).forEach(perform(x){x.classList.take away(‘on’)});t.classList.add(‘on’);$(‘#h2-‘+t.dataset.t).classList.add(‘on’);if(t.dataset.t===’p3′)drawB();if(t.dataset.t===’p4′)drawR();setTimeout(sz,50)}});
/* pipeline */
var nd=$$(‘.nd’),i=-1,tm=null;
var notes=["<b>Reference.</b> The frozen model runs the task with the base harness H0. This shows how the harness behaves, not whether the answer is correct.",
"<b>Propose.</b> An editor drafts 2 candidates: an <b>invariant</b> one for general agent failures (planning, tool use) and a <b>domain</b> one for task knowledge and procedures.",
"<b>Probe.</b> Each candidate runs on the same task. Comparing trajectories against the reference gives a behavioral contrast. No ground-truth labels needed.",
"<b>Compose.</b> The composer reads all evidence and diffs. It may merge, revise, adopt 1 candidate, or keep the current harness.",
"<b>Execute.</b> The composed harness runs and returns the final deliverable.",
"<b>Reflect.</b> The full improvement record is reviewed. Lessons are kept only if execution contrasts support them.",
"<b>Update.</b> The adopted harness and an <code>EXPERIENCE.md</code> of lessons join the improvement library.",
"<b>Next task.</b> The executor resets to H0, but the library now guides proposals and composition. Removing it cut WorkBuddy from 77.5 to 74.4."];
perform present(ok){nd.forEach(perform(n,j){n.classList.toggle(‘lit’,j<=ok)});$(‘#h2-note’).innerHTML=ok<0?"Press <b>Play</b> or <b>Step</b> to walk through 1 Harness<sup>2</sup> improvement step.":notes[k];$(‘#h2-prog’).type.width=((ok+1)/nd.size*100)+’%’;sz()}
perform step(){i=(i+1)%nd.size;present(i)}
$(‘#h2-step’).onclick=perform(){cease();step()};
$(‘#h2-reset’).onclick=perform(){cease();i=-1;present(-1)};
perform cease(){if(tm){clearInterval(tm);tm=null;$(‘#h2-play’).textContent=”▶ Play”}}
$(‘#h2-play’).onclick=perform(){if(tm){cease();return}if(i>=nd.length-1)i=-1;step();tm=setInterval(perform(){if(i>=nd.length-1){cease();return}step()},1900);this.textContent=”❚❚ Pause”};
/* parts */
var C=[["System prompt","Context construction","#4285F4","Persona and top-level procedure followed on every call."],
["Guardrails","Context construction","#4285F4","Process rules, e.g. verify outputs before finishing."],
["Memory","Context construction","#4285F4","Persistent context made available to the agent."],
["Tool guidance","Tool interface","#EA4335","How to use specific tools and handle formats precisely."],
["Subagents","Orchestration","#FBBC04","Delegated roles with their own instructions."],
["Skills","Context construction","#4285F4","Reusable domain procedures, e.g. a reconciliation skill."],
["Scripts","Tool interface","#EA4335","Executable helpers, e.g. a script that reopens and checks files."],
["Plugins","Middleware control","#34A853","Middleware that shapes behavior inside the tool loop."]];
var cg=$(‘#h2-cg’);C.forEach(perform(c,ok){var b=doc.createElement(‘button’);b.className=”cp”;b.type.setProperty(‘–c’,c[2]);b.innerHTML='<em>C’+(ok+1)+’ · ‘+c[1]+'</em><b>’+c[0]+'</b>’;b.onclick=perform(){$$(‘.cp’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);var n=$(‘#h2-cnote’);n.type.borderLeftColor=c[2];n.innerHTML='<b>’+c[0]+'</b> (‘+c[1].toLowerCase()+’): ‘+c[3]+’ Every part is a file the editor can rewrite.’;sz()};cg.appendChild(b)});
/* benchmarks */
var D={"Gemini 3.5 Flash · OpenCode":[[72.4,77.5],[31.5,36.0],[79.9,81.9]],"Gemini 3.5 Flash · Codex":[[74.5,77.5],[33.9,35.3],[71.6,75.6]],"GLM-5 · OpenCode":[[64.2,75.7],[20.6,24.2],[59.6,71.4]],"GLM-5 · Codex":[[64.4,75.3],[23.6,24.8],[58.1,65.9]]};
var BN=["WorkBuddy-Bench (avg)","JobBench","LAB Partial (%)"],cur="GLM-5 · OpenCode";
var sel=$(‘#h2-sel’);Object.keys(D).forEach(perform(ok){var b=doc.createElement(‘button’);b.className=”chip”+(ok===cur?’ on’:”);b.textContent=ok;b.onclick=perform(){cur=ok;$$(‘#h2-sel .chip’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);drawB()};sel.appendChild(b)});
perform drawB(){var bm=$(‘#h2-bm’);bm.innerHTML=”;D[cur].forEach(perform(v,ok){var d=(v[1]-v[0]).toFixed(1);var x=doc.createElement(‘div’);x.className=”bx”;x.innerHTML='<h4>’+BN[k]+'</h4><div class="row"><span>Base</span><div class="tr"><i type="background:#5f6368" data-w="’+v[0]+’"></i></div><b>’+v[0].toFixed(1)+'</b></div><div class="row"><span>Harness²</span><div class="tr"><i type="background:linear-gradient(90deg,#4285F4,#34A853)" data-w="’+v[1]+’"></i></div><b>’+v[1].toFixed(1)+'</b></div><div class="dl">+’+d+(ok===2?’ pp’:’ pts’)+'</div>’;bm.appendChild(x)});
requestAnimationFrame(perform(){setTimeout(perform(){$$(‘#h2-bm .tr i’).forEach(perform(e){e.type.width=e.dataset.w+’%’})},30)});sz()}
/* baselines */
var RK={"WorkBuddy avg · OpenCode":[["Base harness",72.4],["ReasoningBank",73.4],["Harness Scaling",75.3],["Meta-Harness",76.4],["Harness² k=1",77.5],["Harness² k=5",80.8]],
"JobBench · OpenCode":[["Base harness",31.5],["ReasoningBank",30.8],["Harness Scaling",32.9],["Meta-Harness",35.8],["Harness² k=1",36.0]],
"LAB Partial · Codex":[["Base harness",71.6],["ReasoningBank",74.8],["Harness Scaling",67.9],["Meta-Harness",76.5],["Harness² k=1",75.6]]};
var rc="WorkBuddy avg · OpenCode",s2=$(‘#h2-sel2’);
Object.keys(RK).forEach(perform(ok){var b=doc.createElement(‘button’);b.className=”chip”+(ok===rc?’ on’:”);b.textContent=ok;b.onclick=perform(){rc=ok;$$(‘#h2-sel2 .chip’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);drawR()};s2.appendChild(b)});
perform drawR(){var rk=$(‘#h2-rk’);rk.innerHTML=”;var mx=Math.max.apply(null,RK[rc].map(perform(r){return r[1]}));
RK[rc].forEach(perform(r){var h=r[0].indexOf(‘Harness²’)===0;var c=h?’linear-gradient(90deg,#4285F4,#34A853)’:(r[1]===mx?’#FBBC04′:’#5f6368′);var x=doc.createElement(‘div’);x.className=”row”;x.innerHTML='<span’+(h?’ type="color:#e8eaed;font-weight:700"’:”)+’>’+r[0]+'</span><div class="tr"><i type="background:’+c+’" data-w="’+((r[1]-50)/40*100)+’"></i></div><b>’+r[1].toFixed(1)+'</b>’;rk.appendChild(x)});
requestAnimationFrame(perform(){setTimeout(perform(){$$(‘#h2-rk .tr i’).forEach(perform(e){e.type.width=e.dataset.w+’%’})},30)});sz()}
present(-1);window.addEventListener(‘load’,sz);window.addEventListener(‘resize’,sz);setTimeout(sz,300);
})();
</script>
</div>
</physique></html>”>

(perform(){var f=doc.getElementById(‘mtp-h2x-frame’);window.addEventListener(‘message’,perform(e){if(e.supply===f.contentWindow&&e.information&&e.information.mtpH2xHeight){f.type.peak=e.information.mtpH2xHeight+’px’;}});})();

How does Harness² carry out on benchmarks?

The analysis staff examined 3 skilled benchmarks: WorkBuddy-Bench, JobBench and Harvey’s LAB. Executors had been Gemini 3.5 Flash and GLM-5 underneath OpenCode and Codex. The default editor was Claude Opus 4.8. Outcomes are means over 3 runs.

Setup (single step, ok=1) WorkBuddy Avg JobBench LAB Partial
Gemini 3.5 Flash, OpenCode 72.4 → 77.5 31.5 → 36.0 79.9 → 81.9
Gemini 3.5 Flash, Codex 74.5 → 77.5 33.9 → 35.3 71.6 → 75.6
GLM-5, OpenCode 64.2 → 75.7 20.6 → 24.2 59.6 → 71.4
GLM-5, Codex 64.4 → 75.3 23.6 → 24.8 58.1 → 65.9

Supply: Harness² paper, Tables 1 and 2.

Weaker executors gained extra. After 1 step, GLM-5 beat base Gemini 3.5 Flash on WorkBuddy-Bench. Parallel recursion at ok=5 lifted Gemini underneath OpenCode to 80.8 on WorkBuddy. On WorkBuddy-Bench, 3 of 4 setups additionally used fewer steps per process. Gemini underneath OpenCode fell from 25.4 to 21.9.

How does it examine with different harness strategies?

Function Harness² Meta-Harness ReasoningBank Harness Scaling
Builders Google Cloud AI Analysis, Northwestern Stanford IRIS Lab and co-authors Google Analysis and co-authors Wang et al.
What modifications Full 8-component harness, per process Harness code, searched offline Retrieved reminiscence; harness fastened Harness, revised per occasion
Wants coaching break up No Sure No No
Learns throughout duties Sure (library + classes) Deploys 1 finest harness Sure (reminiscence) No
Code license Apache-2.0 MIT Not disclosed Not disclosed
WorkBuddy Avg, Gemini/OpenCode* 77.5 (80.8 at ok=5) 76.4 73.4 75.3
JobBench, Gemini/OpenCode* 36.0 35.8 30.8 32.9
LAB Partial, Gemini/Codex* 75.6 76.5 74.8 67.9

*Scores as measured within the Harness² paper; the opposite strategies didn’t report these benchmarks themselves.

How do you put in and run Harness²?

The README targets Linux, Python 3.12 and uv. A neighborhood demo runs each recursion modes with out mannequin credentials:

Copy CodeCopiedUse a special Browser
uv venv --python 3.12 .venv
supply .venv/bin/activate
uv pip set up -e '.[render]'
harness2-demo --output runs/demo --k 2

Mannequin-backed runs want Node.js 20+, the Google Cloud CLI and a Vertex AI mission. LAB wants Pandoc and Poppler. WorkBuddy-Bench wants Docker. Executors are OpenCode by default, or Codex with --substrate cx.

Key Takeaways

  • Harness² rewrites an agent’s harness at take a look at time; weights keep frozen.
  • 1 step lifted GLM-5 by 11.5 factors on WorkBuddy-Bench.
  • Suggest, Probe, Compose replaces ground-truth labels with behavioral contrasts.
  • A classes library carries enhancements to future duties.
  • Count on additional runs per process, and sandbox it your self.


Try the Paper, GitHub Repo, Paper and Project Page. All credit score goes to the researcher of this mission. Additionally, be happy to observe us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

The put up Google Open-Sources Harness2, a Recursive Agent Harnessing Framework appeared first on MarkTechPost.

ar ces Google work
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min

21/11/2026

Meet Pondering Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac

21/11/2026

Baseten Releases Bison-1 and Groundhog-1, Topping Choice Mannequin Index

21/11/2026

RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Mannequin Skilled With Reinforcement Studying Solely

20/11/2026
Top News

Shut These Laptops! Anthropic Places Its Claude Cowork Agent on Your Cellphone

AI Leaders Call for a Sloweddown. Trump’s Team Claims It Is on Them

AI PACs Have Dumped Almost $1 Million Into an Obscure Senate Race

AI Isn’t Smarter Than a Baby—Yet

How to Use Google Chrome’s New AI-Powered ‘Skills’

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

Gemini Command Line Interface Overview

29/06/2025

The Simple Process to Get Consistent Coaching Clients –

03/09/2025
Latest News

Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains

22/11/2026

Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI

21/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.