NVIDIA Research introduced HORIZONA hands-free framework for designing hardware. It views hardware design in terms of repository-level code. This team of researchers exercises register-transfer-level (RTL)instantiation. An organized Markdown harness is transformed into a project bundle. An isolated git branch is then created by a self-contained agent. It only commits the version when a executable gate is passed.
Researchers report 100% success in every RTL benchmark suite evaluated. The report also makes it clear that the problem of agentic hardware is still not resolved.
What is HORIZON?
The number of design tasks which can be completed in one turn is clearly limited. Real hardware requires more than just plausible Verilog. It is important to consider the simulator’s feedback, cycle-level behavior and reset conventions.
HORIZON stores each design challenge as a repository that is version controlled, and not just a single-shot. A structured Markdown harness is the only input required. This harness has four parts: a goal; domain knowledge directions; an evaluator’s specification and an acceptance condition.
Bootstrap agents compile the harness and project into one package. The research team writes this as p = (πAgentEPAP, ΓP, ΩP). They also cover the version-control policy and domain skills. These terms also include the version control policy and domain skills.
RTL evaluates the RTL by EP Included in this category are compilation, simulations, coverage extractors, assertion or testbench checking, as well as the like. This same slot can be used for unit tests, profiler, theorem provers or synthesis tool in other domains. The problem is defined by git tree worktrees and not a specific repository type.
What is the Repository Level Loop?
The loop will run without any further input from the user after bootstrapping. Every cycle sets a goal, modifies the worktree and invokes tools. It also runs the evaluator. This acceptance condition then determines whether the change is to be committed or logged.
Not incidental bookkeeping, but Git. The diffs reveal proposed changes. Commits define accepted checkpoints. Evidence is attached to notes by the evaluator. The full trajectory is recovered by the log.
This loop relies on native git command to reduce the cost of tracing. Staged edits are inspected with git diff –cached. Accepted attempts become a git-commit whose note contains the result and reward. Commitments that are successful become examples of positive repairs. Rejected efforts are recorded as negative examples. The repository is the buffer for experience, and not a datastore.
For a specific purpose, the research team borrowed semi-Markov decision processes vocabulary. The only thing it does is name the objects that are recorded. A ‘state’ is a versioned snapshot of the repository. A’state’ is a versioned snapshot of the repository. “option” It is only a single episode that separates two checkpoints. HORIZON doesn’t train or update RL policies in this work. The backbone of the agent remains fixed during a campaign.
Session reuse keeps cost down. HORIZON holds a persistent model session across iterations. From the cache, providers can serve up harness, stable sources, and project pack. Tokens that are newly charged will be dominated by both the most recent evaluator and current diff.
HORIZON – A Self-Evolving System
HORIZON continues a long tradition of self-evolution at the repository scale. Early systems evolved software which engineers used. HORIZON, on the other hand, evolves hardware artifacts created by engineers.
| The system is a | Object evolved | The Domain | Evaluation signal |
|---|---|---|---|
| AlphaEvolve (AltaEvolve) (2025). | Algorithmic kernels | Discover scientific and algorithmic discoveries | Automatic evaluators |
| SATLUTION (2020) | Full SAT Solver Repository | SAT-solving | Distributed Correctness and Runtime |
| ABCEvo 2026 | ABC logic-synthesis systems | EDA Software | The QoR and Correctness |
| HORIZON | RTL sources and testbenches. | Hardware design | Compile, simulate, coverage, assertion checks |
They all share a common principle. The candidate for change will only be admitted if it is supported by executable evidence.
Benchmark Results
GPT-5.3 is the backbone of all experiments. All results are in hands-free, single agent mode. The campaigns were run on a 32-core AMD EPYC host, with 512GB RAM.
Verilog Eval, RTLLM-2.0 and ChipBench all contribute to the evaluation. It adds nine CVDP code- and verification-generation categories, CID 002 to 016. CVDP includes 783 manually-written problems in 13 categories of tasks (Pinckney, et. al. 2025).
A iteration can be defined as a step that is automated. The agent then edits and runs the worktree before committing a successful pass or logging a failure. HORIZON achieves a pass rate of 100% on each suite. The one residual miss is a ChipBench specification-harness defect, not an agent failure.
The first-iteration aggregate pass rate is 48.8%. Iteration-0 does not stand alone as a Pass@1 measure. The repository is in the same state as after the agent’s first iteration. Debugging and repairs may be deferred to later agent iterations.
| Category / Suite | Focus | Iter. 0 | Conv. iter. | HORIZON |
|---|---|---|---|---|
| ChipBench | Mixed RTL generation | 20.0 | 5 | 100.0 |
| RTLLM-2.0 | RTL spec to NL | 78.0 | 2 | 100.0 |
| Verilog-Eval-v2 | HDLBits-style Verilog | 86.2 | 2 | 100.0 |
| CVDP CID 0002 | RTL Code Completion | 3.2 | 82 | 100.0 |
| CVDP CID CID 003 | RTL spec to NL | 19.2 | 24 | 100.0 |
| CVDP CID CID 004 | RTL Code Modification | 10.9 | 36 | 100.0 |
| CVDP CID CID005 | Spec-to-RTL module reuse | 9.1 | 14 | 100.0 |
| CVDP CID CID007 | Linting/QoR Improvement | 0.0 | 24 | 100.0 |
| CVDP CID 012 | Test plan for stimuli generation | 47.8 | 32 | 100.0 |
| CVDP CID 013 | Test plan to checker generator | 3.8 | 19 | 100.0 |
| CVDP CID 014 | Test plan to assertion generation | 79.1 | 1 | 100.0 |
| CVDP CID 016 | Bug fixing and debugging | 25.7 | 13 | 100.0 |
Convergence difficulties vary widely between categories. RTLLM and Verilog Eval both reach 100 % in just two iterations. The checker generation (CID 013) begins at only 3.8%. It is still climbing steadily, reaching 100% after 19 iterations. The code completion for CID 002 requires 82 iterations. The cost of the tokens is highest in its long tail.
The Interactive Metrics Explainer
‘+
‘
‘+
‘
‘
‘;
tr.onclick=function(){[].forEach.call(rb.children,function(x){x.classList.remove(‘sel’)});tr.classList.add(‘sel’)};
rb.appendChild(tr);
});
TABS
[].forEach.call(root.querySelectorAll(‘.tab’),function(tab){
tab.onclick=function(){
root.querySelectorAll(‘.tab’).forEach(function(x){x.classList.remove(‘on’)});
root.querySelectorAll(‘.pane’).forEach(function(x){x.classList.remove(‘on’)});
tab.classList.add(‘on’);document.getElementById(tab.dataset.t).classList.add(‘on’);
setTimeout(resize,60);
};
});
// LOOP
STEPS=[
{t:’Task Action Plan’,d:’The agent reads the current worktree state and plans a target edit.’},
{t:’File Edit & Tool Calls’,d:’It edits candidate artifacts (RTL, testbench) and invokes domain tools.’},
{t:’Evaluate and Score’,d:’The executable evaluator Ep runs compile / simulate / coverage and emits evidence yt.’},
{t:’Correctness Gate & Review’,d:’Acceptance predicate Ap checks yt; an independent step diffs the candidate.’},
{t:’Commit & Git Trace’,d:’On pass: git commit with git notes carrying the verdict and reward. On fail: RejectLog as a negative example.’}
];
var ls=document.getElementById(‘loopSteps’),lstate=document.getElementById(‘loopState’),cur=-1;
function drawLoop(){
ls.innerHTML=”;
STEPS.forEach(function(s,i){
var el=document.createElement(‘div’);el.className=”step”+(i===cur?’ act’:”);
el.innerHTML=’
‘+(i+1)+’
‘;
ls.appendChild(el);
});
}
drawLoop();
document.getElementById(‘loopNext’).onclick=function(){
cur++;
if(cur>=STEPS.length){cur=0;var it=(+lstate.textContent.replace(/D/g,”))+1;lstate.innerHTML=’State Sw,’+it+’ (committed)’;}
drawLoop();setTimeout(resize,60);
};
document.getElementById(‘loopReset’).onclick=function(){cur=-1;lstate.innerHTML=’State Sw,0‘;drawLoop();setTimeout(resize,60);};
CONVERGENCE
var sel=document.getElementById(‘convSel’);
S.forEach(function(d,i){var o=document.createElement(‘option’);o.value=i;o.textContent=d.n+’ u2014 ‘+d.f;sel.appendChild(o);});
Sel.value=3 // CID 0002 default value (the longtail)
var timer=null;
function ease(x){return 1-Math.pow(1-x,2.2);}
function runConv(){
if(timer)clearInterval(timer);
var d=S[sel.value],max=Math.max(d.c,1);
document.getElementById(‘convMax’).textContent=max;
var log=document.getElementById(‘convLog’);log.innerHTML=”;
If it is 0, then the value of it will be 0.
var step=Math.max(1,Math.round(max/40));
timer=setInterval(function(){
it+=step; if(it>max)it=max;
It is a maximum value.
var pass=d.i0+(100-d.i0)*ease(frac);
var tok=d.t*frac;
document.getElementById(‘convPass’).textContent=pass.toFixed(1);
document.getElementById(‘convPassBar’).style.width=pass+’%’;
document.getElementById(‘convTok’).textContent=tok.toFixed(1);
document.getElementById(‘convIter’).textContent=it;
var pass2=Math.random();
if(it===max){log.insertAdjacentHTML(‘afterbegin’,’
iter ‘+it+’ u2713 pass=100% u00b7 gate satisfied
‘);}
else if(pass2>0.5){log.insertAdjacentHTML(‘afterbegin’,’
iter ‘+it+’ u2713 partial repair accepted
‘);}
else{log.insertAdjacentHTML(‘afterbegin’,’
reject-log iter ‘+it+’ u2717 evaluator mismatch (negative example)
‘);}
Resize();
if(it>=max)clearInterval(timer);
},110);
}
document.getElementById(‘convRun’).onclick=runConv;
document.getElementById(‘convReset’).onclick=function(){
if(timer)clearInterval(timer);
document.getElementById(‘convPass’).textContent=”0.0″;document.getElementById(‘convPassBar’).style.width=”0%”;
document.getElementById(‘convTok’).textContent=”0.0″;document.getElementById(‘convIter’).textContent=”0″;
document.getElementById(‘convLog’).innerHTML=”;resize();
};
DON’T
Var groups=[{n:’Legacy suites’,v:2.9,c:’#5a6b3f’},{n:’CID 002′,v:26.7,c:’#76B900′},{n:’CID 003′,v:18.1,c:’#9ede3a’},{n:’CID 004′,v:11.3,c:’#557a1a’},{n:’CID 005′,v:4.4,c:’#b6e86a’},{n:’CID 007′,v:10.3,c:’#6ea015′},{n:’CID 012′,v:15.3,c:’#8fce2e’},{n:’CID 013′,v:6.7,c:’#3f5c12′},{n:’CID 014′,v:0.1,c:’#cdd7bd’},{n:’CID 016′,v:4.2,c:’#456218′}];
var svg=document.getElementById(‘donut’),leg=document.getElementById(‘donutLeg’),ang=-90,cx=60,cy=60,r=44,rr=27;
groups.forEach(function(g){
var a0=ang,a1=ang+g.v/100*360;ang=a1;
var la=(a1-a0)>180?1:0;
function pt(a,rad){var rd=a*Math.PI/180;return [cx+rad*Math.cos(rd),cy+rad*Math.sin(rd)];}
var p0=pt(a0,r),p1=pt(a1,r),p2=pt(a1,rr),p3=pt(a0,rr);
The path is:”M”+p0[0]+’ ‘+p0[1]+’A’+r+’ ‘+r+’ 0 ‘+la+’ 1 ‘+p1[0]+’ ‘+p1[1]+’L’+p2[0]+’ ‘+p2[1]+’A’+rr+’ ‘+rr+’ 0 ‘+la+’ 0 ‘+p3[0]+’ ‘+p3[1]+’Z’;
var el=document.createElementNS(‘http://www.w3.org/2000/svg’,’path’);el.setAttribute(‘d’,path);el.setAttribute(‘fill’,g.c);svg.appendChild(el);
var sp=document.createElement(‘span’);sp.innerHTML=’‘+g.n+’ ‘+g.v+’%’;leg.appendChild(sp);
});
var ctr=document.createElementNS(‘http://www.w3.org/2000/svg’,’text’);ctr.setAttribute(‘x’,60);ctr.setAttribute(‘y’,58);ctr.setAttribute(’text-anchor’,’middle’);ctr.setAttribute(‘font-size’,’11’);ctr.setAttribute(‘font-weight’,’800′);ctr.textContent=”210M”;svg.appendChild(ctr);
var ctr2=document.createElementNS(‘http://www.w3.org/2000/svg’,’text’);ctr2.setAttribute(‘x’,60);ctr2.setAttribute(‘y’,70);ctr2.setAttribute(’text-anchor’,’middle’);ctr2.setAttribute(‘font-size’,’7′);ctr2.setAttribute(‘fill’,’#9aa691′);ctr2.textContent=”tokens”;svg.appendChild(ctr2);
// RESIZE
Resize the image using this function(){var h=root.offsetHeight;parent.postMessage({type:’mtpHorizonResize’,height:h+40},’*’);}
window.addEventListener(‘load’,function(){setTimeout(resize,80);});
window.addEventListener(‘resize’,resize);
setTimeout(resize,200);
})();

