Meet Manufacturing facility Droid 1.0: A Mannequin-Impartial Coding Agent Harness Constructed on Script Mode

by root November 1, 2026

written by root November 1, 2026 0 comment 1 views

Manufacturing facility has launched Droid 1.0, the most important replace but to the Manufacturing facility Droid agent harness. The primary change is programmatic device calling, now on by default. The mannequin writes a program that runs its instruments. One mannequin request can now drive tons of of device calls.

TL;DR

Dimension: Base immediate is 13,055 enter tokens, down 37% from 20,887. Solely 2 device definitions ship up entrance, down from 11.
Runs on: macOS and Home windows desktop app, a terminal UI, and the Node SDK @manufacturing unit/droid-sdk. Works with any supported mannequin, together with your personal API keys.
Efficiency: Highest go fee of each harness Manufacturing facility examined, on the identical mannequin, at decrease price.
Greatest: 65.2% on Terminal-Bench 4 with Claude Opus 5.5, at $4.73 per handed process (self-reported).
Worst: With GPT-6.1 Sol, Droid takes 10% longer per handed process than Codex.

What’s Droid 1.0?

Droid 1.0 is a model-independent coding agent harness from Manufacturing facility. Manufacturing facility defines a harness as every thing across the mannequin. Which means the prompts, the tools, how tool results return, and the loop. Customers choose a mannequin or let Factory Router select a cost-efficient combine.

How does programmatic device calling work?

In a regular loop, each device output lands within the mannequin’s context. With programmatic device calling (PTC), the mannequin writes a program as a substitute. Droid implements this as a Script device. The mannequin writes JavaScript that runs in a sandbox. The sandbox has no file, community, or system entry of its personal.

Uncooked recordsdata, search outcomes, and logs keep inside this system. Solely printed output returns to the mannequin. Impartial calls run concurrently.

The thought builds on CodeAct (Wang et al., 2024). Manufacturing facility stories these outcomes from its personal classes:

Unit check repair: 15 device calls took 5 mannequin calls as a substitute of 10. The duty completed in 41.6 s as a substitute of 56.8 s.
Linear cleanup: 151 points moved and verified in 312 device calls. The Script used 2 requests and about 61,000 tokens. Manufacturing facility fashions direct calls at 36 requests and about 3.1M tokens.
CI sweep: 5,384 jobs throughout 547 runs, in 2 requests. Underneath 0.1% of device outcomes reached the mannequin’s context.

On Claude Opus 5.5, Script mode cuts Droid’s price by 15% versus direct device calling. It’s on by default for 10 fashions, together with GPT-6.1 Sol and Kimi K3.

<div class="pn on" id="p1">
<h3>Direct calls vs the Script device</h3>
<p>Fixing a failing unit check with GPT-6.1 Sol: the identical 15 device calls, replayed each methods (Manufacturing facility's measurements).</p>
<button class="btn" id="fdxRun"> Play the race</button>
<div class="race">
<div class="lane"><b>Direct device calls</b><div class="dots" id="lnA"></div><div class="stat"><span>Mannequin calls <robust id="aM">0</robust></span><span>Instrument calls <robust id="aT">0</robust></span><span>Time <robust id="aS">0.0 s</robust></span></div></div>
<div class="lane"><b>Script device</b><div class="dots" id="lnB"></div><div class="stat"><span>Mannequin calls <robust id="bM">0</robust></span><span>Instrument calls <robust id="bT">0</robust></span><span>Time <robust id="bS">0.0 s</robust></span></div></div>
</div>
<div class="lg"><span><i type="background:#2f6bff"></i>Mannequin name</span><span><i type="background:#c9c9c9"></i>Instrument name</span><span><i type="background:#111"></i>Script</span></div>
<p class="be aware">The Script is JavaScript in a sandbox with no file, community or system entry of its personal. Uncooked device outcomes keep in this system; solely printed output returns to the mannequin.</p>
</div>

<div class="pn" id="p2">
<h3>Terminal-Bench 4 go fee, identical mannequin per harness</h3>
<div class="seg"><button class="on" data-m="opus">Claude Opus 5.5</button><button data-m="sol">GPT-6.1 Sol</button></div>
<div id="fdxBars"></div>
<p class="be aware" id="fdxCost"></p>
<p class="be aware">Self-reported by Manufacturing facility. 3 trials per process, run October 23 to twenty-eight, 2026. Cursor was not examined on GPT-6.1 Sol.</p>
</div>

<div class="pn" id="p3">
<h3>Shifting 151 Linear points: 312 device calls</h3>
<p>Every situation was moved to the backlog, then checked in order that solely its standing modified.</p>
<button class="btn gh" id="fdxLin"> Evaluate</button>
<div class="cmp" type="margin-top:12px">
<div class="field"><div class="lbl">Direct calls (modeled, 10 per request)</div><div class="massive" id="lA">0</div><div class="lbl">mannequin requests</div><div class="massive" id="lAt" type="font-size:22px;margin-top:8px">0</div><div class="lbl">enter tokens</div><div class="tk"><div id="lAb" type="background:#9aa6bf"></div></div></div>
<div class="field"><div class="lbl">One Script</div><div class="massive b" id="lB">0</div><div class="lbl">mannequin requests</div><div class="massive b" id="lBt" type="font-size:22px;margin-top:8px">0</div><div class="lbl">enter tokens</div><div class="tk"><div id="lBb" type="background:#2f6bff"></div></div></div>
</div>
<p class="be aware">Direct-call figures are Manufacturing facility’s estimates; the Script figures come from an actual session.</p>
</div>

<div class="pn" id="p4">
<h3>Different adjustments in Droid 1.0 (faucet a card)</h3>
<div class="playing cards">
<div class="cd"><div class="n">−37%</div><div class="t">Base immediate</div><div class="x">First request fell from 20,887 to 13,055 enter tokens. Droid now sends 2 device definitions up entrance as a substitute of 11.</div></div>
<div class="cd"><div class="n">98%</div><div class="t">Cache reads</div><div class="x">A typical manufacturing request reads 98% of its enter from cache. Newly loaded instruments are appended on the finish so the cached prefix holds.</div></div>
<div class="cd"><div class="n">−15%</div><div class="t">Script mode price</div><div class="x">Script mode cuts Droid’s price by 15% on Claude Opus 5.5 in contrast with direct device calling.</div></div>
<div class="cd"><div class="n">16K</div><div class="t">Compaction room</div><div class="x">Compaction summaries can now maintain 16,000 tokens, up from 8,000. 3 equivalent device calls set off a nudge; 6 cease the loop.</div></div>
<div class="cd"><div class="n">488 ms</div><div class="t">First paint</div><div class="x">Median time to first paint fell from 1,839 ms to 488 ms. Median stop time fell from 648 ms to 30 ms.</div></div>
<div class="cd"><div class="n">~20 MB</div><div class="t">Per further session</div><div class="x">Periods now share the daemon course of, including about 20 MB every as a substitute of about 350 MB per course of.</div></div>
</div>
</div>

<div class="ft"><span>Supply: <a href="https://manufacturing unit.com/information/frontier-harness" goal="_blank" rel="noopener">Manufacturing facility Analysis, Oct 29, 2026</a> · figures self-reported</span><span class="mtp">© Marktechpost</span></div>
</div>
<script>
(operate(){
var W=doc.getElementById(‘fdx-wrap’);
operate rs(){attempt{mother or father.postMessage({fdxHeight:W.offsetHeight+40},’*’)}catch(e){}}
W.querySelectorAll(‘.tab’).forEach(operate
/* race */
var A=[[1,2],[1,1],[1,1],[1,1],[1,1],[1,2],[1,2],[1,2],[1,2],[1,1]],B=[[1,6],[1,2],[1,2],[1,4],[1,1]];
operate race(){var la=doc.getElementById(‘lnA’),lb=doc.getElementById(‘lnB’);la.innerHTML=”;lb.innerHTML=”;
operate seq(steps,lane,ids,whole,scr){var objects=[];steps.forEach(operate(s,i){objects.push(‘m’);var final=(i===steps.length-1);if(scr&&!final){objects.push(‘s’)}for(var ok=0;ok<s[1];ok++)objects.push(‘t’)});
var m=0,t=0,i=0,step=whole/objects.size;var iv=setInterval(operate(){if(i>=objects.size){clearInterval(iv);doc.getElementById(ids[2]).textContent=whole.toFixed(1)+’ s’;return}var d=doc.createElement(‘span’);d.className=”d “+objects[i];lane.appendChild(d);setTimeout(operate(){d.classList.add(‘present’)},20);if(objects[i]===’m’)m++;if(objects[i]===’t’)t++;doc.getElementById(ids[0]).textContent=m;doc.getElementById(ids[1]).textContent=t;doc.getElementById(ids[2]).textContent=(step*(i+1)).toFixed(1)+’ s’;i++;rs()},step*45)}
seq(A,la,[‘aM’,’aT’,’aS’],56.8,false);seq(B,lb,[‘bM’,’bT’,’bS’],41.6,true)}
doc.getElementById(‘fdxRun’).onclick=race;
/* bars */
var D={opus:{rows:[[‘Droid’,65.2],[‘Cursor’,62.1],[‘OpenCode’,62.1],[‘Devin’,59.6],[‘Claude Code’,57.6]],c:’Droid: $4.73 per handed process, as much as 24% cheaper and 16% to 26% quicker than each different harness on Opus 5.5.’},sol:{rows:[[‘Droid’,57.6],[‘Codex’,54.0],[‘OpenCode’,51.0],[‘Devin’,45.0]],c:’Droid: $0.79 per handed process, 12% cheaper than Codex however 10% slower per handed process.’}},cur=”opus”;
operate bars(ok){cur=ok;var h=””;D[k].rows.forEach(operate(r){h+='<div class="bar"><span>’+r[0]+'</span><div class="trk"><div class="fill’+(r[0]===’Droid’?’ me’:”)+’" data-w="’+(r[1]/70*100)+’"></div></div><span class="val">’+r[1].toFixed(1)+’%</span></div>’});var b=doc.getElementById(‘fdxBars’);b.innerHTML=h;doc.getElementById(‘fdxCost’).textContent=D[k].c;setTimeout(operate(){b.querySelectorAll(‘.fill’).forEach(operate(f){f.type.width=f.dataset.w+’%’})},40);rs()}
W.querySelectorAll(‘.seg button’).forEach(operate(b){b.onclick=operate(){W.querySelectorAll(‘.seg button’).forEach(operate(x){x.classList.take away(‘on’)});b.classList.add(‘on’);bars(b.dataset.m)}});
/* linear */
operate cnt(el,to,fmt,ms){var s=null;operate f(ts){if(!s)s=ts;var p=Math.min((ts-s)/ms,1);el.textContent=fmt(to*p);if(p<1)requestAnimationFrame(f)}requestAnimationFrame(f)}
doc.getElementById(‘fdxLin’).onclick=operate(){cnt(doc.getElementById(‘lA’),36,operate(v){return Math.spherical(v)},1100);cnt(doc.getElementById(‘lB’),2,operate(v){return Math.spherical(v)},600);cnt(doc.getElementById(‘lAt’),3.1,operate(v){return ‘~’+v.toFixed(1)+’M’},1100);cnt(doc.getElementById(‘lBt’),61,operate(v){return ‘~’+Math.spherical(v)+’Ok’},600);doc.getElementById(‘lAb’).type.width=”100%”;doc.getElementById(‘lBb’).type.width=(61/3100*100).toFixed(2)+’%’};
W.querySelectorAll(‘.cd’).forEach(operate(c){c.onclick=operate(){c.classList.toggle(‘open’);setTimeout(rs,380)}});
bars(‘opus’);window.addEventListener(‘load’,rs);window.addEventListener(‘resize’,rs);setTimeout(rs,300);
})();
</script>
</physique></html>
“>

window.addEventListener(‘message’,operate(e){if(e.knowledge&&e.knowledge.fdxHeight){var f=doc.getElementById(‘fdx-frame’);if(f)f.type.top=e.knowledge.fdxHeight+’px’;}});

How does Droid 1.0 rating on Terminal-Bench 4?

Manufacturing facility ran every harness towards Droid on the identical mannequin. Every harness ran 3 trials per process, from October 23 to twenty-eight, 2026. Solely duties each harnesses handed in all 3 trials rely towards price and time.

Claude Opus 5.5: Droid scores 65.2%. Cursor and OpenCode rating 62.1%, Devin 59.6%, Claude Code 57.6%. Droid prices as much as 24% much less.
GPT-6.1 Sol: Droid scores 57.6% at $0.79 per handed process. Codex scores 54.0%, OpenCode 51.0%, Devin 45.0%.
Pace: On Opus 5.5, Droid takes 16% to 26% much less time per handed process. In opposition to Claude Code, it additionally makes use of 17% fewer tokens.

What else modified within the harness?

Most adjustments got here from an inside Agent Manufacturing facility loop. Brokers learn failures and suggest adjustments, examined on each affected mannequin household. Engineers evaluation and merge.

Immediate cache: Newly loaded instruments are appended to the request finish. The standard manufacturing request reads 98% of enter from cache.
Leaner first request: Instrument definitions fell 46%. Hardly ever used instruments now load on demand.
Bounded autonomy: 3 equivalent device calls set off a nudge, and 6 cease the loop. Compaction summaries now maintain 16,000 tokens, up from 8,000.
App latency: First paint fell from 1,839 ms to 488 ms. Every further session now provides about 20 MB, not 350 MB.

Comparability: Droid 1.0 vs different coding agent harnesses

Characteristic	Droid 1.0 (Factory)	Codex CLI	Claude Code	OpenCode
Developer	Manufacturing facility	OpenAI	Anthropic	Anomaly
Mannequin assist	Mannequin-independent, BYOK, Factory Router	ChatGPT plan or API key	Native harness for Claude (Factory)	75+ LLM providers, local models
License	Not disclosed (SDK is Apache 2.0)	Apache-2.0	Anthropic Commercial Terms	MIT
Programmatic device calling	Default (Script device)	Sure, per Factory	Not disclosed	Not disclosed
TB4, Claude Opus 5.5*	65.2%	Not examined	57.6%	62.1%
TB4, GPT-6.1 Sol*	57.6%	54.0%	Not examined	51.0%
Time per handed process vs Droid*	Baseline	Droid 10% slower (Sol)	Droid 26% quicker (Opus)	Droid 4% to 24% quicker
Pricing	From $20/mo	ChatGPT plan or API	Not disclosed	Free, open supply

*Terminal-Bench 4 figures are Manufacturing facility’s personal runs, from its methodology section.

How can builders attempt it?

Droid 1.0 ships as a desktop app and terminal UI. The Droid SDK runs the identical harness from Node, with classes, streaming and MCP instruments. Paid plans begin at $20 per month.

Key Takeaways

Droid 1.0 makes programmatic device calling the default.
1 Script can run tons of of device calls per mannequin request.
65.2% on Terminal-Bench 4 with Opus 5.5, self-reported.
Base immediate shrank 37%, to 13,055 tokens.
Codex nonetheless wins on velocity with GPT-6.1 Sol.

Take a look at the total Factory research post, the Droid SDK on GitHub and the Droid docs. Additionally, be happy to comply with us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Have to associate with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and many others.? Connect with us

The publish Meet Manufacturing facility Droid 1.0: A Mannequin-Impartial Coding Agent Harness Constructed on Script Mode appeared first on MarkTechPost.

Welcome to Ivugangingo!

At Ivugangingo, we're passionate about delivering insightful content that empowers and informs our readers across a spectrum of crucial topics. Whether you're delving into the world of insurance, navigating the complexities of cryptocurrency, or seeking wellness tips in health and fitness, we've got you covered.

Meet Manufacturing facility Droid 1.0: A Mannequin-Impartial Coding Agent Harness Constructed on Script Mode

TL;DR

What’s Droid 1.0?

How does programmatic device calling work?

How does Droid 1.0 rating on Terminal-Bench 4?

What else modified within the harness?

Comparability: Droid 1.0 vs different coding agent harnesses

How can builders attempt it?

Key Takeaways

Sturdy Serialization with JSON and Customized Sorts

Which is true to your crew?

Converter

Editors Pick

Newsletter

Categories

Related Posts

Leave a Comment Cancel Reply