An AI search agent can work in a demo however fail below manufacturing workloads when retrieval returns incomplete proof, retries repeat unsuccessful actions, or execution exceeds time and price limits. TinyFish, which gives internet infrastructure for AI brokers, has revealed two technical guides addressing these challenges. The primary defines a production agentic search architecture, whereas the second shares lessons from building a deep research system. Collectively, they look at how an autonomous search agent can plan retrieval, consider proof, deal with failures, and handle the boundary between mannequin reasoning and deterministic execution.
TL;DR
TinyFish launched its agentic search structure and key engineering classes from constructing deep analysis programs. Constructed on 4 internet primitives (TinySearch, TinyFetch, TinyBrowser, and TinyAgent), the system operates by way of a seven-phase management loop across the base mannequin. On OpenBenchmarks, TinyFish achieved a 79.0% completion rate on coding duties whereas delivering the bottom median activity value ($0.050) and token footprint, highlighting the significance of inflexible management loops, failure-specific retries, and unified reasoning over standalone planners.
What’s Agentic Search Structure?
Agentic search architecture allows an AI agent to plan retrieval, entry sources, consider proof, and adapt its method as new data turns into obtainable. The mannequin gives reasoning, whereas the encircling utility manages state, instrument execution, proof dealing with, runtime insurance policies, and output validation. Collectively, these parts permit an agent to deal with incomplete outcomes, recoverable failures, and altering data necessities.
TinyFish’s structure attracts foundational ideas from a February 2026 architecture preprint, which decouples reasoning from execution by typed instrument interfaces. The paper establishes important system guardrails, together with detailed tracing alongside strict caps on tokens, execution time, general value, and gear invocations.
How does a manufacturing search run work?
TinyFish splits a run into 7 phases:
- Outline the duty: Flip the objective into an “proof contract”: required fields, sources, freshness and completion guidelines.
- Plan retrieval: Separate unbiased wants from dependent ones. Dependent queries anticipate earlier solutions.
- Choose the primary route: Select a instrument primarily based on what the system is aware of in regards to the proof.
- Uncover, choose, retrieve: Deal with search outcomes as candidates till the web page content material arrives.
- Consider: Label proof as ample, incomplete, conflicting, inaccessible or price range exhausted.
- Get better or escalate: Revise the question, retry, change sources or change entry methodology.
- Cease and validate: Return accomplished, incomplete, failed, cancelled or human assessment required.
Part 2 addresses a identified failure. ChainRAG (ACL 2025) calls it the ‘lost-in-retrieval’ downside, the place key entities disappear throughout multi-hop query decomposition. Part 6 attracts on Search-R1, which demonstrates how fashions can study to generate extra search queries throughout reasoning when earlier retrieval is inadequate. Throughout 7 QA datasets, its realized multi-turn search beat RAG baselines by 41% (Qwen2.5-7B) and 20% (Qwen2.5-3B).
How ought to a search agent route instruments?
Routing ought to observe the knowledge hole, not a set ladder.
| Beginning situation | Primitive | What the app nonetheless owns |
| Supply location unknown | Search | Question planning, supply coverage, stopping |
| URL identified and readable | Fetch | Completeness checks, declare validation |
| App should management web page actions | Browser | Each motion, retry and checkpoint |
| Choices throughout a number of pages | Web Agent | Enterprise guidelines, budgets, validation |
Escalation will not be at all times the repair. Weak outcomes want a greater question. Conflicting sources want stronger proof. A heavier instrument fixes neither.
Retries must be primarily based on the failure sort and stay inside outlined limits. Transient errors might justify repeating a request, whereas incomplete or conflicting proof requires a special method. Browser retries should additionally account for actions already accomplished to keep away from unintended penalties, comparable to submitting the identical type twice.
<html lang="en"><head><meta charset="utf-8"><meta title="viewport" content material="width=device-width,initial-scale=1">
<title>Agentic Search Explainer</title>
<fashion>
html,physique{margin:0!vital;padding:0!vital;background:clear!vital}
#mtp-tfas{–bg:#111;–card:#161616;–line:#2c2c2c;–g:#76B900;–t:#eaeaea;–m:#8f8f8f;–amb:#e0a03a;–red:#e05a4a;
max-width:880px;margin:0 auto;background:var(–bg)!vital;coloration:var(–t)!vital;border:1px stable var(–line)!vital;border-radius:14px;
font:15px/1.45 system-ui,-apple-system,"Segoe UI",Roboto,sans-serif;overflow:hidden;box-sizing:border-box}
#mtp-tfas *{box-sizing:border-box}
#mtp-tfas hr,#mtp-tfas p:empty,#mtp-tfas del,#mtp-tfas s{show:none!vital}
#mtp-tfas .hd{padding:18px 22px 4px;show:flex;align-items:baseline;hole:12px;flex-wrap:wrap}
#mtp-tfas .kick{coloration:var(–g)!vital;font-size:12px;letter-spacing:.12em;text-transform:uppercase;font-weight:700}
#mtp-tfas h3{margin:0!vital;font-size:20px;coloration:#fff!vital;line-height:1.25;flex-basis:100%}
#mtp-tfas .tabs{show:flex;hole:6px;padding:12px 22px 0;border-bottom:1px stable var(–line);overflow-x:auto}
#mtp-tfas .tab{background:clear!vital;border:1px stable var(–line)!vital;border-bottom:none!vital;coloration:var(–m)!vital;
padding:9px 14px;border-radius:9px 9px 0 0;cursor:pointer;font:600 13px system-ui,sans-serif;white-space:nowrap}
#mtp-tfas .tab.on{background:var(–card)!vital;coloration:var(–g)!vital;border-color:var(–g)!vital}
#mtp-tfas .pane{show:none;padding:14px 22px 16px;background:var(–card)!vital}
#mtp-tfas .pane.on{show:block;animation:tfIn .35s ease}
@keyframes tfIn{from{opacity:0;remodel:translateY(6px)}to{opacity:1;remodel:none}}
#mtp-tfas .chips{show:flex;flex-wrap:wrap;hole:6px;justify-content:middle}
#mtp-tfas .chip{background:#1d1d1d!vital;border:1px stable var(–line)!vital;coloration:var(–t)!vital;padding:7px 12px;border-radius:20px;cursor:pointer;font:600 13px system-ui,sans-serif}
#mtp-tfas .chip.on{background:var(–g)!vital;coloration:#111!vital;border-color:var(–g)!vital}
#mtp-tfas svg{show:block;width:100%;max-width:700px;peak:auto;margin:6px auto 0}
#mtp-tfas svg textual content{font-family:system-ui,-apple-system,"Segoe UI",Roboto,sans-serif}
#mtp-tfas .cap{text-align:middle;coloration:var(–m)!vital;font-size:13px;margin:4px 0 0;min-height:19px}
#mtp-tfas .cap b{coloration:var(–g)!vital}
#mtp-tfas .cap a,#mtp-tfas a{coloration:var(–g)!vital}
#mtp-tfas .ring{stroke-dasharray:6 10;animation:tfFlow 1.2s linear infinite}
@keyframes tfFlow{to{stroke-dashoffset:-16}}
#mtp-tfas .wire{stroke-dasharray:7 7;animation:tfFlow .7s linear infinite}
#mtp-tfas .ft{show:flex;justify-content:space-between;hole:10px;padding:10px 22px;border-top:1px stable var(–line);font-size:12px;coloration:var(–m)!vital;background:var(–bg)!vital;flex-wrap:wrap}
#mtp-tfas .ft .model{coloration:var(–g)!vital;font-weight:700}
@media (max-width:640px){
#mtp-tfas .hd{padding:14px 14px 2px}#mtp-tfas .tabs{padding:10px 14px 0}#mtp-tfas .pane{padding:12px 10px 14px}
#mtp-tfas h3{font-size:17px}#mtp-tfas .tab{padding:8px 10px;font-size:12px}#mtp-tfas .chip{padding:6px 10px;font-size:12px}#mtp-tfas .ft{padding:10px 14px}
}
</fashion></head><physique>
<div id="mtp-tfas">
<div class="hd"><span class="kick">Explainer · Agentic search</span><h3>How a manufacturing search agent works</h3></div>
<div class="tabs" function="tablist">
<button class="tab on" data-t="p1">The loop</button>
<button class="tab" data-t="p2">Instrument routing</button>
<button class="tab" data-t="p3">Deep analysis</button>
<button class="tab" data-t="p4">Benchmark</button>
</div>
<div class="pane on" id="p1">
<div class="chips" id="c1">
<button class="chip" data-s="clear">Clear run</button>
<button class="chip on" data-s="weak">Weak outcomes</button>
<button class="chip" data-s="price range">Funds hit</button>
</div>
<svg id="s1" viewBox="0 0 640 350" aria-label="Agentic search management loop"></svg>
<p class="cap" id="k1"></p>
</div>
<div class="pane" id="p2">
<div class="chips" id="c2">
<button class="chip on" data-r="0">Supply unknown</button>
<button class="chip" data-r="1">URL identified</button>
<button class="chip" data-r="2">Code drives the web page</button>
<button class="chip" data-r="3">Multi-page objective</button>
<button class="chip" data-r="4">Weak or conflicting</button>
</div>
<svg id="s2" viewBox="0 0 640 320" aria-label="Instrument routing"></svg>
<p class="cap" id="k2"></p>
</div>
<div class="pane" id="p3">
<div class="chips" id="c3">
<button class="chip on" data-v="0">v1 Planner</button>
<button class="chip" data-v="1">v2 Subagents</button>
<button class="chip" data-v="2">v3 Instrument calls</button>
<button class="chip" data-v="3">Supply weight</button>
</div>
<svg id="s3" viewBox="0 0 640 300" aria-label="Deep analysis structure variations"></svg>
<p class="cap" id="k3"></p>
</div>
<div class="pane" id="p4">
<div class="chips" id="c4">
<button class="chip on" data-b="sf">Search + fetch</button>
<button class="chip" data-b="so">Search solely</button>
</div>
<svg id="s4" viewBox="0 0 640 340" aria-label="Completion versus value scatter"></svg>
<p class="cap">Median activity value vs grounded completion, 100 duties, gpt-5.6-sol. Supply: <a href="https://openbenchmarks.com/web-search-for-coding-agents" goal="_blank" rel="noopener">OpenBenchmarks</a>, 26 Sept 2026.</p>
</div>
<div class="ft"><span>Sources: <a href="https://www.tinyfish.ai/weblog/production-agentic-search-architecture" goal="_blank" rel="noopener">structure</a> · <a href="https://www.tinyfish.ai/weblog/lessons-from-building-a-deep-research-system" goal="_blank" rel="noopener">deep analysis</a> · <a href="https://openbenchmarks.com/web-search-for-coding-agents" goal="_blank" rel="noopener">OpenBenchmarks</a></span><span class="model">Constructed by Marktechpost</span></div>
</div>
<script>
(operate(){
var R=doc.getElementById('mtp-tfas'),NS='http://www.w3.org/2000/svg';
var G='#76B900',LN='#2c2c2c',BX='#1d1d1d',TX='#eaeaea',MU='#8f8f8f',AM='#e0a03a',RD='#e05a4a';
operate $(s){return R.querySelector(s)}operate $$(s){return R.querySelectorAll(s)}
operate el(t,a,p){var e=doc.createElementNS(NS,t);for(var ok in a)e.setAttribute(ok,a[k]);if(p)p.appendChild(e);return e}
operate tx(p,x,y,s,a){var o={x:x,y:y,fill:TX,'font-size':15,'text-anchor':'center','dominant-baseline':'center'};for(var ok in a)o[k]=a[k];var t=el('textual content',o,p);t.textContent=s;return t}
operate clear(s){whereas(s.firstChild)s.removeChild(s.firstChild)}
operate enhance(sv,f){if(R.offsetWidth>=560)return;sv.querySelectorAll('textual content').forEach(operate
operate put up(){attempt{mum or dad.postMessage({mtpTfasH:R.offsetHeight+40},'*')}catch(e){}}
var energetic='p1';
$$('.tab').forEach(operate
operate chips(id,attr,fn){$$('#'+id+' .chip').forEach(operate(c){c.onclick=operate(){$$('#'+id+' .chip').forEach(operate(x){x.classList.take away('on')});c.classList.add('on');fn(c.dataset[attr])}})}
operate arrowDefs(s,id,col){var d=el('defs',{},s);var m=el('marker',{id:id,viewBox:'0 0 10 10',refX:9,refY:5,markerWidth:7,markerHeight:7,orient:'auto-start-reverse'},d);el('path',{d:'M0 0L10 5L0 10z',fill:col},m)}
/* ———- TAB 1: the loop ———- */
var s1=$('#s1'),C={x:330,y:180,r:115};
var N={objective:[80,52],plan:[200,52],route:[330,65],retrieve:[445,180],consider:[330,295],recuperate:[215,180],validate:[545,295],partial:[110,295]};
var LB={objective:'Objective',plan:'Plan',route:'Route',retrieve:'Retrieve',consider:'Consider',recuperate:'Get better',validate:'Validate',partial:'Partial'};
var ANG={route:-90,retrieve:0,consider:90,recuperate:180};
var field={},packet,bud,budT,center1,center2,termTxt;
operate build1(){clear(s1);arrowDefs(s1,'a1',LN);
el('circle',{cx:C.x,cy:C.y,r:C.r,fill:'none',stroke:'#3a3a3a','stroke-width':2.5,class:'ring'},s1);
[['goal','plan'],['plan','route']].forEach(operate(p){var a=N[p[0]],b=N[p[1]];el('line',{x1:a[0]+52,y1:a[1],x2:b[0]-54,y2:b[1]>a[1]?b[1]:a[1],stroke:LN,'stroke-width':2,'marker-end':'url(#a1)'},s1)});
el('line',{x1:382,y1:295,x2:491,y2:295,stroke:LN,'stroke-width':2,'marker-end':'url(#a1)'},s1);
el('line',{x1:278,y1:295,x2:164,y2:295,stroke:LN,'stroke-width':2,'marker-end':'url(#a1)'},s1);
for(var ok in N){var g=el('g',{},s1),p=N[k];var r=el('rect',{x:p[0]-52,y:p[1]-19,width:104,peak:38,rx:10,fill:BX,stroke:LN,'stroke-width':1.5},g);
var t=tx(g,p[0],p[1]+1,LB[k],{'font-size':15,'font-weight':600});field[k]={r:r,t:t}}
center1=tx(s1,C.x,C.y-10,'',{'font-size':22,'font-weight':800,fill:'#fff'});
center2=tx(s1,C.x,C.y+18,'',{'font-size':14,fill:MU});
tx(s1,430,24,'Process price range',{'font-size':12,fill:MU,'text-anchor':'begin'});
el('rect',{x:430,y:34,width:180,peak:10,rx:5,fill:'#222'},s1);
bud=el('rect',{x:430,y:34,width:0,peak:10,rx:5,fill:G},s1);
budT=tx(s1,612,24,'0%',{'font-size':12,fill:TX,'text-anchor':'finish'});
tx(s1,330,340,'time · tokens · instrument calls · retries · value',{'font-size':11,fill:'#5c5c5c'});
packet=el('circle',{r:8,fill:G,stroke:'#111','stroke-width':2},s1);
el('circle',{r:16,fill:'none',stroke:G,'stroke-opacity':.35,'stroke-width':2,id:'pkh'},s1);enhance(s1,1.3)}
var SC={
clear:{steps:[['goal'],['plan'],['route'],['retrieve'],['evaluate','Sufficient',G],['validate','Completed',G]],cap:0.42,txt:'Proof meets the contract on the primary go. <b>Cease and validate.</b>'},
weak:{steps:[['goal'],['plan'],['route'],['retrieve'],['evaluate','Incomplete',AM],['recover','Rewrite query',AM],['route'],['retrieve'],['evaluate','Sufficient',G],['validate','Completed',G]],cap:0.74,txt:'A spot sends the run again across the loop with a <b>modified</b> question.'},
price range:{steps:[['goal'],['plan'],['route'],['retrieve'],['evaluate','Incomplete',AM],['recover','Retry, bounded',AM],['route'],['retrieve'],['evaluate','Budget exhausted',RD],['partial','Partial + review',AM]],cap:1,txt:'On the boundary the agent <b>stops truthfully</b>, not with a fluent guess.'}};
operate seg(a,b){var pts=[],i;
if(ANG[a]!==undefined&&ANG[b]!==undefined){var a0=ANG[a],a1=ANG[b];if(a1<=a0)a1+=360;for(i=0;i<=40;i++){var an=(a0+(a1-a0)*i/40)*Math.PI/180;pts.push([C.x+C.r*Math.cos(an),C.y+C.r*Math.sin(an)])}}
else{var p=N[a],q=N[b];for(i=0;i<=30;i++)pts.push([p[0]+(q[0]-p[0])*i/30,p[1]+(q[1]-p[1])*i/30])}
return pts}
var run1=null;
operate play1(title){var sc=SC[name];$('#k1').innerHTML=sc.txt;
var path=[],marks=[];for(var i=0;i<sc.steps.size;i++){if(i>0){var sg=seg(sc.steps[i-1][0],sc.steps[i][0]);sg.shift();path=path.concat(sg)}else path.push(N[sc.steps[0][0]].slice());marks.push(path.length-1)}
var L=[0];for(i=1;i<path.size;i++)L.push(L[i-1]+Math.hypot(path[i][0]-path[i-1][0],path[i][1]-path[i-1][1]));var tot=L[L.length-1];
for(var ok in field){field[k].r.setAttribute('stroke',LN);field[k].r.setAttribute('fill',BX);field[k].t.textContent=LB[k];field[k].t.setAttribute('fill',TX)}
var velocity=0.16,maintain=650,finished=-1,pk=$('#pkh'),dist=0,prev=null,wait=0,ending=false;
operate reset(){dist=0;finished=-1;wait=0;ending=false;for(var ok in field){field[k].r.setAttribute('stroke',LN);field[k].r.setAttribute('fill',BX);field[k].t.textContent=LB[k];field[k].t.setAttribute('fill',TX)}center1.textContent='';center2.textContent=''}
reset();
run1={tick:operate(ts){if(prev===null)prev=ts;var dt=ts-prev;prev=ts;if(dt>100)dt=16;
if(wait>0){wait-=dt;if(wait<=0&&ending){reset();return}}else dist=Math.min(tot,dist+dt*velocity);
var d=dist;
var nxt=finished+1;if(nxt<marks.size&&d>=L[marks[nxt]]){finished=nxt;dist=d=L[marks[nxt]];wait=maintain;if(finished===marks.length-1){wait=1900;ending=true}var st=sc.steps[done],b=field[st[0]];
b.r.setAttribute('stroke',st[2]||G);b.r.setAttribute('fill',st[2]===RD?'#2a1412':st[2]===AM?'#2a2111':'#1b2410');
center1.textContent=LB[st[0]];center2.textContent=st[1]||'';center2.setAttribute('fill',st[2]||MU);
if(st[1]&&(st[0]==='validate'||st[0]==='partial')){b.t.textContent=st[1];b.t.setAttribute('fill',st[2])}
}
var j=0;whereas(j<L.length-1&&L[j+1]<d)j++;var f=L[j+1]>L[j]?(d-L[j])/(L[j+1]-L[j]):0;f=Math.max(0,Math.min(1,f||0));
var p=j<path.length-1?[path[j][0]+(path[j+1][0]-path[j][0])*f,path[j][1]+(path[j+1][1]-path[j][1])*f]:path[path.length-1];
packet.setAttribute('cx',p[0]);packet.setAttribute('cy',p[1]);pk.setAttribute('cx',p[0]);pk.setAttribute('cy',p[1]);pk.setAttribute('r',12+4*Math.sin(ts/150));
var fr=d/tot*sc.cap;bud.setAttribute('width',180*fr);bud.setAttribute('fill',fr>=.99?RD:fr>.6?AM:G);budT.textContent=Math.spherical(fr*100)+'%'}}}
build1();chips('c1','s',play1);play1('weak');
/* ———- TAB 2: routing ———- */
var s2=$('#s2'),lanes=[],wire,qmark,needT,rw;
var LA=[['Search','find candidate sources','srch'],['Fetch','clean page content','doc'],['Browser','your code drives it','win'],['Web Agent','goal in, JSON out (Mako)','bot']];
operate icon(g,ok,x,y){var st={fill:'none',stroke:'currentColor','stroke-width':2.2,'stroke-linecap':'spherical'};
if(ok==='srch'){el('circle',Object.assign({cx:x-3,cy:y-3,r:8},st),g);el('line',Object.assign({x1:x+3,y1:y+3,x2:x+9,y2:y+9},st),g)}
if(ok==='doc'){el('rect',Object.assign({x:x-8,y:y-10,width:16,peak:20,rx:2},st),g);[ -4,0,4].forEach(operate(o){el('line',Object.assign({x1:x-4,y1:y+o,x2:x+4,y2:y+o},st),g)})}
if(ok==='win'){el('rect',Object.assign({x:x-11,y:y-9,width:22,peak:18,rx:2},st),g);el('line',Object.assign({x1:x-11,y1:y-4,x2:x+11,y2:y-4},st),g)}
if(ok==='bot'){el('rect',Object.assign({x:x-9,y:y-6,width:18,peak:14,rx:4},st),g);el('line',Object.assign({x1:x,y1:y-6,x2:x,y2:y-11},st),g);el('circle',{cx:x-4,cy:y+1,r:1.8,fill:'currentColor'},g);el('circle',{cx:x+4,cy:y+1,r:1.8,fill:'currentColor'},g)}}
operate build2(){clear(s2);arrowDefs(s2,'a2',G);
tx(s2,110,40,'Proof want',{'font-size':12,fill:MU});
el('circle',{cx:110,cy:150,r:52,fill:'#1b2410',stroke:G,'stroke-width':2},s2);
qmark=tx(s2,110,150,'?',{'font-size':40,'font-weight':800,fill:G});
needT=tx(s2,110,222,'',{'font-size':13,fill:TX});
wire=el('path',{fill:'none',stroke:G,'stroke-width':2.5,class:'wire','marker-end':'url(#a2)'},s2);
LA.forEach(operate(l,i){var y=42+i*70,g=el('g',{fashion:'coloration:'+MU},s2);
var r=el('rect',{x:380,y:y-24,width:240,peak:52,rx:11,fill:BX,stroke:LN,'stroke-width':1.5},g);icon(g,l[2],410,y+2);
var t=tx(g,436,y-6,l[0],{'font-size':17,'font-weight':700,'text-anchor':'begin'});tx(g,436,y+13,l[1],{'font-size':12,fill:MU,'text-anchor':'begin'});
lanes.push({g:g,r:r,t:t,y:y})});
rw=el('g',{opacity:0},s2);el('rect',{x:30,y:262,width:160,peak:44,rx:10,fill:'#2a2111',stroke:AM,'stroke-width':1.5},rw);
tx(rw,110,284,'Rewrite question',{'font-size':15,'font-weight':700,fill:AM});enhance(s2,1.22)}
var RT=[['no URL yet',0,'Discover before paying to inspect or operate pages.'],['have a URL',1,'Skip search. It cannot add discovery value here.'],['scripted clicks',2,'Deterministic navigation stays in your code.'],['many pages',3,'Keeps the goal and run state while the path changes.'],['bad evidence',-1,'<b>No escalation.</b> A heavier tool will not fix a weak query.']];
operate route(i){i=+i;var r=RT[i];needT.textContent=r[0];$('#k2').innerHTML=r[2];
lanes.forEach(operate(l,ok){var on=ok===r[1];l.r.setAttribute('stroke',on?G:LN);l.r.setAttribute('fill',on?'#1b2410':BX);l.g.setAttribute('fashion','coloration:'+(on?G:MU)+';opacity:'+(r[1]<0?.35:on?1:.55));l.t.setAttribute('fill',on?'#fff':TX)});
if(r[1]>=0){var y=lanes[r[1]].y;wire.setAttribute('d','M164 150 C 270 150, 270 '+y+', 374 '+y);wire.setAttribute('opacity',1);rw.setAttribute('opacity',0);qmark.textContent='?'}
else{wire.setAttribute('d','M110 204 L 110 256');rw.setAttribute('opacity',1);qmark.textContent='!'}
wire.animate&&wire.animate([{opacity:0},{opacity:1}],{period:400})}
build2();chips('c2','r',route);route(0);
/* ———- TAB 3: deep analysis ———- */
var s3=$('#s3'),dots=[],spawn=null,stage=0,last3=0;
operate tagPill(x,y,s,c){var w=s.size*7.4+26;el('rect',{x:x-w/2,y:y-14,width:w,peak:28,rx:14,fill:c==G?'#1b2410':c==RD?'#2a1412':'#2a2111',stroke:c,'stroke-width':1.5},s3);tx(s3,x,y+1,s,{'font-size':13,'font-weight':700,fill:c})}
operate bx(x,y,w,h,label,col,fs){el('rect',,s3);if(label)tx(s3,x+w/2,y+h/2,label,15,'font-weight':700)}
var weight=0;
operate v3(v){v=+v;stage=v;clear(s3);dots=[];arrowDefs(s3,'a3',LN);
if(v===0){bx(30,50,240,200,'',LN);tx(s3,150,72,'Orchestrator',{'font-size':15,'font-weight':700});
var bars=[['structured state',G],['chat history','#5b8def'],['past reports','#5b8def'],['findings','#5b8def'],['user intent','#5b8def']];
bars.forEach(operate(b,i){var y=96+i*30;el('rect',{x:50,y:y,width:200,peak:22,rx:5,fill:b[1],'fill-opacity':.18,stroke:b[1],'stroke-opacity':.6},s3);tx(s3,62,y+11,b[0],{'font-size':12,'text-anchor':'begin',fill:TX})});
el('line',{x1:340,y1:40,x2:340,y2:260,stroke:RD,'stroke-width':2,'stroke-dasharray':'6 6'},s3);tx(s3,340,30,'boundary',{'font-size':11,fill:RD});
bx(420,110,190,80,'',LN);tx(s3,515,140,'Planner',{'font-size':16,'font-weight':700});tx(s3,515,164,'contemporary LLM name',{'font-size':12,fill:MU});
tagPill(515,236,'Context misplaced',RD);
spawn=operate(){var i=Math.flooring(Math.random()*5),y=107+i*30;dots.push({x:250,y:y,tx:420,ty:150,c:i?'#5b8def':G,die:i?340:null,v:2.4})};
$('#k3').innerHTML='Solely structured state crossed. Observe-ups like "focus solely on pricing dangers" drifted.'}
if(v===1){bx(30,110,170,80,'Orchestrator',RD);var ys=[50,150,250];
ys.forEach(operate(y,i){bx(430,y-26,180,52,'Subagent '+(i+1),LN,14);el('line',{x1:200,y1:150,x2:430,y2:y,stroke:LN,'stroke-width':1.5},s3)});
tagPill(115,240,'Latency ↑ Value ↑',RD);
spawn=operate(){var y=ys[Math.floor(Math.random()*3)],out=Math.random()<.5;dots.push(out?{x:200,y:150,tx:430,ty:y,c:MU,v:4}:{x:430,y:y,tx:200,ty:150,c:AM,v:4})};
$('#k3').innerHTML='1 subagent per question. The orchestrator saved dispatching, ready and merging.'}
if(v===2){bx(30,100,190,100,'',G);tx(s3,125,135,'Orchestrator',{'font-size':16,'font-weight':700});tx(s3,125,160,'plan + resolve subsequent',{'font-size':12,fill:G});
['Search','Fetch','Browse'].forEach(operate(n,i){var y=60+i*60;bx(290,y-20,110,40,n,LN,14);el('line',{x1:220,y1:150,x2:290,y2:y,stroke:LN,'stroke-width':1.5},s3);el('line',{x1:400,y1:y,x2:460,y2:120,stroke:LN,'stroke-width':1.5},s3)});
bx(460,70,150,100,'',G);tx(s3,535,106,'Rank + extract',{'font-size':14,'font-weight':700});tx(s3,535,128,'plain code',{'font-size':12,fill:G});
el('path',{d:'M535 170 C 535 250, 125 260, 125 204','fill':'none',stroke:G,'stroke-width':1.5,'stroke-dasharray':'4 5','marker-end':'url(#a3)'},s3);
tx(s3,330,262,'ranked sources + extracted claims',{'font-size':12,fill:G});
tagPill(330,20,'Context ↓ Latency ↓ Value ↓',G);
spawn=operate(){var i=Math.flooring(Math.random()*3),y=60+i*60;dots.push({x:400,y:y,tx:460,ty:120,c:MU,v:2.2,r:6,then:1})};
$('#k3').innerHTML='Deterministic steps turned instrument calls. The orchestrator solely sees the distilled end result.'}
if(v===3){spawn=null;
var w=el('g',{},s3),gl=[['Goal: product sentiment',0],['Goal: unemployment rate',1]];
gl.forEach(operate(g,i){var x=i?330:30,on=weight===i,rg=el('g',{fashion:'cursor:pointer'},w);el('rect',{x:x,y:10,width:280,peak:34,rx:17,fill:on?G:BX,stroke:on?G:LN},rg);tx(rg,x+140,28,g[0],{'font-size':13,'font-weight':700,fill:on?'#111':TX});rg.onclick=operate(){weight=i;v3(3)}});
tx(s3,170,272,'Social posts',{'font-size':14,'font-weight':700});tx(s3,480,272,'Authorities report',{'font-size':14,'font-weight':700});
var soc=[[110,110],[160,90],[210,120],[130,170],[185,165],[235,180],[95,215],[150,225],[215,225]];
soc.forEach(operate(p){var c=el('circle',{cx:p[0],cy:p[1],r:6,fill:weight===0?G:'#3a3a3a'},s3);c.animate&&c.animate([{r:6},{r:weight===0?20:6}],{period:700,fill:'forwards',easing:'ease-out'});if(weight===0)c.setAttribute('r',20)});
var gv=el('circle',{cx:480,cy:160,r:weight===1?78:16,fill:weight===1?G:'#3a3a3a','fill-opacity':.9},s3);gv.animate&&gv.animate([{r:16},{r:weight===1?78:16}],{period:700,easing:'ease-out'});
$('#k3').innerHTML='Similar sources, completely different weight. <b>Rank by the analysis goal</b>, not one high quality rating.'}
enhance(s3,1.25);setTimeout(put up,60)}
operate tick3(ts){if(spawn&&ts-last3>(stage===1?110:260)){spawn();last3=ts}
for(var i=dots.length-1;i>=0;i–){var d=dots[i];if(!d.e){d.e=el('circle',,s3)}
var dx=d.tx-d.x,dy=d.ty-d.y,l=Math.hypot(dx,dy);
if(d.die&&d.x>=d.die-2){d.e.setAttribute('fill',RD);d.fade=(d.fade||1)-.06;d.e.setAttribute('opacity',d.fade);if(d.fade<=0){d.e.take away();dots.splice(i,1)}proceed}
if(l<d.v){if(d.then){d.e.take away();dots.splice(i,1);dots.push({x:535,y:170,tx:125,ty:204,c:G,v:3.2,r:3.5,curve:1});proceed}d.e.take away();dots.splice(i,1);proceed}
d.x+=dx/l*d.v;d.y+=dy/l*d.v;d.e.setAttribute('cx',d.x);d.e.setAttribute('cy',d.y)}}
chips('c3','v',v3);v3(0);
/* ———- TAB 4: scatter ———- */
var s4=$('#s4'),D={sf:[['TinyFish',0.050,79.0],['Exa deep',0.127,83.0],['Perplexity',0.083,77.7],['Firecrawl',0.082,76.0]],
so:[['TinyFish',0.035,59.3],['Exa fast',0.103,66.3],['Perplexity',0.059,77.3],['Firecrawl',0.059,70.3]]};
var X0=80,X1=610,Y0=290,Y1=30,pts={},cur='sf';
operate sx(v){return X0+(v/0.14)*(X1-X0)}operate sy(v){return Y0-((v-50)/40)*(Y0-Y1)}
operate build4(){clear(s4);
for(var c=0;c<=0.14001;c+=0.02){var x=sx(c);el('line',{x1:x,y1:Y1,x2:x,y2:Y0,stroke:'#222'},s4);tx(s4,x,Y0+18,'$'+c.toFixed(2),{'font-size':12,fill:MU})}
for(var v=50;v<=90;v+=10){var y=sy(v);el('line',{x1:X0,y1:y,x2:X1,y2:y,stroke:'#222'},s4);tx(s4,X0-10,y,v+'%',{'font-size':12,fill:MU,'text-anchor':'finish'})}
tx(s4,(X0+X1)/2,Y0+40,'Median activity value (decrease is healthier)',{'font-size':12,fill:MU});
tx(s4,24,(Y0+Y1)/2,'Process completion',{'font-size':12,fill:MU,remodel:'rotate(-90 24 '+(Y0+Y1)/2+')'});
tx(s4,X0+12,Y1+14,' higher’,{‘font-size’:13,’font-weight’:700,fill:G,’text-anchor’:’begin’});
D.sf.forEach(operate(d){var tf=d[0]===’TinyFish’,g=el(‘g’,{},s4);
var ring=el(‘circle’,{r:tf?17:0,fill:’none’,stroke:G,’stroke-opacity’:.4,’stroke-width’:2},g);
var c=el(‘circle’,{r:tf?10:8,fill:tf?G:’#6a6a6a’,stroke:’#111′,’stroke-width’:2},g);
var t=tx(g,0,0,”,{‘font-size’:13,’font-weight’:700,fill:tf?G:TX,’text-anchor’:’begin’});
pts[d[0]]={g:g,c:c,t:t,ring:ring,x:sx(d[1]),y:sy(d[2])}});
pts[‘Exa fast’]=pts[‘Exa deep’];enhance(s4,1.2)}
var anim4=null,OFF={sf:{‘TinyFish’:[-16,-20,’end’],’Exa deep’:[-14,-18,’end’],’Perplexity’:[14,-12,’start’],’Firecrawl’:[14,16,’start’]},
so:{‘TinyFish’:[16,0,’start’],’Exa quick’:[-14,0,’end’],’Perplexity’:[14,0,’start’],’Firecrawl’:[14,0,’start’]}};
operate drawScatter(fromZero){var information=D[cur],begin={},t0=null;
information.forEach(operate(d){var p=pts[d[0]];begin[d[0]]=fromZero?{x:sx(0),y:Y0}:{x:p.x,y:p.y};p.tx=sx(d[1]);p.ty=sy(d[2]);
p.t.textContent=d[0]+’ ‘+d[2].toFixed(1)+’% · $’+d[1].toFixed(3)});
anim4=operate(ts){if(!t0)t0=ts;var f=Math.min(1,(ts-t0)/900),e=1-Math.pow(1-f,3);
information.forEach(operate(d){var p=pts[d[0]],s=begin[d[0]];p.x=s.x+(p.tx-s.x)*e;p.y=s.y+(p.ty-s.y)*e;p.g.setAttribute(‘remodel’,’translate(‘+p.x+’,’+p.y+’)’);
var o=OFF[cur][d[0]];p.t.setAttribute(‘x’,o[0]);p.t.setAttribute(‘y’,o[1]);p.t.setAttribute(‘text-anchor’,o[2])});
if(f>=1)anim4=null}}
build4();chips(‘c4′,’b’,operate(b){cur=b;drawScatter(false)});drawScatter(true);
operate loop(ts){if(energetic===’p1’&&run1)run1.tick(ts);if(energetic===’p3′)tick3(ts);if(anim4)anim4(ts);
if(pts.TinyFish)pts.TinyFish.ring.setAttribute(‘r’,15+3*Math.sin(ts/220));requestAnimationFrame(loop)}
requestAnimationFrame(loop);
window.addEventListener(‘load’,put up);window.addEventListener(‘resize’,put up);setTimeout(put up,300);
})();
</script>
</physique></html>
“>
window.addEventListener(“message”,operate(e){if(e.information&&e.information.mtpTfasH){var f=doc.getElementById(“mtp-tfas-frame”);if(f)f.fashion.peak=e.information.mtpTfasH+”px”;}});
What did TinyFish study constructing deep analysis?
The deep research agent searches, browses, and generates complete, cited stories. Most engineering challenges centered round context administration, system orchestration, and product design.
Lesson #1: Eradicating the separate planner
Preliminary iterations used a standalone planner known as contemporary every time. Solely structured state crossed into it. On follow-up queries like “focus solely on pricing dangers in Southeast Asia,” the agent drifted again to broad market analysis. TinyFish subsequently moved planning into the orchestrator’s personal unified reasoning loop.
Lesson #2: Eradicating subagents:
TinyFish initially assigned a separate subagent to every search question, rising context-switching overhead, latency, and price. The workforce later changed these subagents with instrument requires repeatable duties comparable to rating search outcomes, classifying sources, and extracting proof. This decreased the knowledge handed again to the orchestrator, permitting it to concentrate on analysis planning and deciding what to research subsequent.
Lesson #3- Matching researchers:
TinyFish initially optimized its analysis system for supply protection, however customers positioned larger significance on the standard and relevance of supporting proof. The workforce adjusted supply rating to account for the analysis goal and added controls for prioritizing, excluding, or proscribing domains. It additionally developed an inner analysis harness utilizing an LLM choose to evaluate stories for relevance, proof power, quotation belief, and supply range.
How do value and latency add up?
TinyFish budgets the total activity, not single calls:
- Process latency = planning + search + retrieval + analysis + retries + interactive steps + queueing.
- Process value = tokens + instrument calls + web page or browser utilization + retries + infrastructure.
SearchAgent-X (ICLR 2026) exhibits why. Precise retrieval added overhead, and coarse retrieval compelled additional reasoning. Its fixes reached as much as 3.4x throughput and 5x decrease latency versus examined baselines.
How does TinyFish evaluate on search and fetch?
OpenBenchmarks fixes the mannequin (gpt-5.6-sol), 100 duties and budgets, and varies solely the supplier. Scores are means of three runs.
| Metric | TinyFish | Exa deep | Perplexity | Firecrawl |
| Search + fetch completion | 79.0% ± 2.0 | 83.0% ± 1.0 | 77.7% ± 1.5 | 76.0% ± 1.0 |
| Median activity time | 24s | 37s | 22s | 34s |
| Median activity value | $0.050 | $0.127 | $0.083 | $0.082 |
| Median activity tokens | 12,844 | 23,660 | 20,062 | 17,379 |
| Search-only completion | 59.3% | 66.3% (Exa quick) | 77.3% | 70.3% |
| Search listing worth | $0 | $0.012 | $0.005 | $0.005 |
Supply: OpenBenchmarks board, up to date 26 Sept 2026. Duties cowl enterprise documentation solely.
OpenBenchmarks additionally stories search-only outcomes individually. In that configuration, TinyFish achieved 59.3% activity completion, in contrast with 77.3% for Perplexity, 70.3% for Firecrawl, and 66.3% for Exa Quick. These outcomes shouldn’t be immediately in contrast with the search-and-fetch scores as a result of the obtainable retrieval capabilities differ.
Key Takeaways
- Management the entire trajectory, not 1 search name.
- Route by data hole; escalate solely when entry fails.
- Preserve planning and execution in 1 reasoning loop.
- Transfer deterministic steps into instrument calls, not subagents.
- Rank sources by analysis goal, not one high quality rating.
Learn the architecture guide, the deep research lessons and the TinyFish docs.
Because of the Tinyfish workforce for the thought management / assets for this text. This text is sponsored by TinyFish.
The put up TinyFish Breaks Down Agentic Search Structure for Manufacturing AI Brokers appeared first on MarkTechPost.
