Sakana AI が企業向けエージェント「Sakana Marlin」を商用化、最大100ページの調査報告書とスライドを生成

Tokyo-based Sakana AI shipped its first commercial product ‘Sakana Marlin’ this week. Sakana team positions it as a Virtual CSO (Chief Strategy Officer). It is a B2B autonomous research agent built for enterprises.

Marlin does not answer in seconds like a chatbot. You give it one research topic. It then runs autonomously for up to about eight hours. Each run returns a long report plus a presentation slide deck. Sakana says a single session issues hundreds to thousands of LLM queries.

What is Sakana Marlin

Marlin is an enterprise research agent, not a chat assistant. You give it one topic or question. It then plans hypotheses, browses sources, and verifies findings on its own. It compresses weeks of strategy work into hours.

The deliverable is structured for decision-makers. The Japanese announcement describes reports of dozens of pages. The English announcement cites reports of up to roughly 100 pages. At a press hands-on, reports ran 60–100 pages and cited 60–80 sources. Each report includes a main body, references, and appendices. Presentation slides are generated using image-generation AI.

Sakana team refined Marlin through a closed beta in April 2026. Around 300 professionals tested it on real tasks during that beta. Those tasks spanned strategy formulation, market research, risk analysis, and competitive analysis. Sakana has also partnered with MUFG and taken strategic investment from Citigroup.

Inside AB-MCTS: Wider or Deeper

The backbone of Marlin is AB-MCTS, or Adaptive Branching Monte Carlo Tree Search. It comes from the Sakana’s past research “Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search.”

AB-MCTS treats reasoning as a tree-search problem. At each step the algorithm makes one decision. It can go wider by generating a new candidate answer. Or it can go deeper by refining a promising existing answer. Standard repeated sampling only goes wider in parallel, then hopes one answer is right.

A multi-LLM variant adds a second choice. It can route a step to a different model entirely. In Sakana’s reported ARC-AGI-2 experiments, this collaboration helped. Combining o4-mini, Gemini 2.5 Pro, and DeepSeek-R1 solved about 27.5% of tasks. The o4-mini model alone solved about 23%. Marlin applies the same adaptive search to long-horizon research.

The second key component for Marlin is workflow automation from Sakana’s AI Scientist project. That project demonstrated autonomous scientific discovery and was published in Nature.

Interactive demo: The embeddable widget (marlin-abmcts-demo.html) shows the “wider or deeper” decision live. Press Run and watch the tree grow. Greener nodes carry higher scores, and the best path is highlighted. Toggle “Multi-LLM” to see steps routed across different models.

#mtp-abmcts-demo *{box-sizing:border-box!important;margin:0;padding:0}

#mtp-abmcts-demo{

background:#111!important;color:#e8e8e8!important;border:1px solid #2a2a2a!important;

border-radius:14px!important;max-width:880px;margin:24px auto;padding:20px 20px 14px;

font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;

line-height:1.5;overflow:hidden

}

#mtp-abmcts-demo h3{color:#fff!important;font-size:19px;font-weight:700;margin:0 0 4px}

#mtp-abmcts-demo .mtp-sub{color:#9aa0a6!important;font-size:13px;margin-bottom:14px}

#mtp-abmcts-demo .mtp-controls{display:flex;flex-wrap:wrap;gap:8px;align-items:center;margin-bottom:12px}

#mtp-abmcts-demo button{

background:#76B900!important;color:#0a0a0a!important;border:0!important;border-radius:8px!important;

font-weight:700;font-size:13px;padding:9px 14px;cursor:pointer;transition:filter .15s

}

#mtp-abmcts-demo button:hover{filter:brightness(1.1)}

#mtp-abmcts-demo button.mtp-ghost{background:#1c1c1c!important;color:#76B900!important;border:1px solid #2f3b1a!important}

#mtp-abmcts-demo button:disabled{opacity:.4;cursor:not-allowed}

#mtp-abmcts-demo .mtp-slider{display:flex;align-items:center;gap:8px;font-size:12px;color:#bdbdbd!important;margin-left:auto}

#mtp-abmcts-demo input[type=range]{accent-color:#76B900;width:120px}

#mtp-abmcts-demo .mtp-toggle{display:flex;align-items:center;gap:6px;font-size:12px;color:#bdbdbd!important;cursor:pointer;user-select:none}

#mtp-abmcts-demo .mtp-stage{

background:#0b0b0b!important;border:1px solid #232323!important;border-radius:10px!important;

position:relative;overflow:hidden;min-height:260px

}

#mtp-abmcts-demo svg{display:block;width:100%;height:auto}

#mtp-abmcts-demo .mtp-edge{stroke:#3a3a3a;stroke-width:1.5;fill:none}

#mtp-abmcts-demo .mtp-edge.best{stroke:#76B900;stroke-width:2.6}

#mtp-abmcts-demo text{font-family:inherit}

#mtp-abmcts-demo .mtp-grid{display:grid;grid-template-columns:1fr 1fr;gap:12px;margin-top:12px}

#mtp-abmcts-demo .mtp-panel{background:#0b0b0b!important;border:1px solid #232323!important;border-radius:10px!important;padding:12px}

#mtp-abmcts-demo .mtp-panel h4{color:#76B900!important;font-size:12px;letter-spacing:.05em;text-transform:uppercase;margin-bottom:8px}

#mtp-abmcts-demo .mtp-stat{display:flex;justify-content:space-between;font-size:13px;padding:3px 0;border-bottom:1px solid #1a1a1a}

#mtp-abmcts-demo .mtp-stat b{color:#fff!important}

#mtp-abmcts-demo .mtp-log{font-size:12.5px;max-height:132px;overflow-y:auto}

#mtp-abmcts-demo .mtp-log div{padding:3px 0;border-bottom:1px solid #1a1a1a;color:#cfcfcf!important}

#mtp-abmcts-demo .mtp-log .w{color:#7fb3ff!important}

#mtp-abmcts-demo .mtp-log .d{color:#76B900!important}

#mtp-abmcts-demo .mtp-legend{display:flex;flex-wrap:wrap;gap:14px;font-size:11.5px;color:#9aa0a6!important;margin-top:10px;align-items:center}

#mtp-abmcts-demo .mtp-chip{display:inline-flex;align-items:center;gap:5px}

#mtp-abmcts-demo .mtp-dot{width:10px;height:10px;border-radius:50%;display:inline-block}

#mtp-abmcts-demo .mtp-foot{text-align:center;color:#76B900!important;font-size:12px;font-weight:600;margin-top:14px;padding-top:10px;border-top:1px solid #232323}

#mtp-abmcts-demo hr,#mtp-abmcts-demo p:empty,#mtp-abmcts-demo del,#mtp-abmcts-demo s{display:none!important}

@media (max-width:640px){

#mtp-abmcts-demo{padding:14px}

#mtp-abmcts-demo .mtp-grid{grid-template-columns:1fr}

#mtp-abmcts-demo .mtp-slider{margin-left:0;width:100%}

#mtp-abmcts-demo input[type=range]{flex:1}

}

AB-MCTS: “Wider or Deeper?” — interactive search

A simplified visual of Sakana AI’s Adaptive Branching Monte Carlo Tree Search. Each step the policy chooses to widen (new candidate) or deepen (refine a promising line).

► Run

Step

Reset

Multi-LLM

DeeperWider

Search state

Budget used0 / 24

Nodes (candidates)1

Best score0.00

Wider / Deeper0 / 0

Decision log

low score

high score

best path

Gemini 2.5 Pro

o4-mini

DeepSeek-R1

© Marktechpost · Illustrative model of AB-MCTS (TreeQuest, Apache 2.0)

(function(){

var SVGNS="http://www.w3.org/2000/svg";

var BUDGET=24, MODELS=[{n:"Gemini 2.5 Pro",c:"#7fb3ff"},{n:"o4-mini",c:"#ffb454"},{n:"DeepSeek-R1",c:"#ff7a90"}];

var nodes, idc, used, widers, deepers, timer=null;

var svg=document.getElementById("mtp-svg");

var $=function(id){return document.getElementById(id);};

function reset(){

if(timer){clearInterval(timer);timer=null;$("mtp-run").textContent="► Run";}

idc=0; used=0; widers=0; deepers=0;

nodes=[{id:idc++,parent:null,depth:0,score:0,model:0,children:[]}];

$("mtp-log").innerHTML="";

log('Initialized search tree at the root.','');

render();

}

function chooseModel(){return Math.floor(Math.random()*MODELS.length);}

// One AB-MCTS step: descend from root, at each node decide widen-here vs go-deeper.

function step(){

if(used>=BUDGET) return false;

var bias=(+$("mtp-bias").value)/100; // 1 = always wider, 0 = always deeper

var cur=nodes[0];

// Descend until we decide to branch (widen) at the current node.

while(true){

var kids=cur.children.map(function(i){return nodes[i];});

// Probability of branching here (going wider) rises with the bias and with few kids.

var pWiden = bias*0.6 + 0.4/(kids.length+1);

if(kids.length===0 || Math.random()40) box.removeChild(box.lastChild);

}

// Tidy-tree layout: leaves get sequential x, parents centered over children.

function layout(){

var leafX=0, pos={};

(function walk(n){

if(n.children.length===0){ pos[n.id]={x:leafX++,y:n.depth}; return; }

var xs=[];

n.children.forEach(function(i){ walk(nodes[i]); xs.push(pos[nodes[i].id].x); });

pos[n.id]={x:(Math.min.apply(null,xs)+Math.max.apply(null,xs))/2,y:n.depth};

})(nodes[0]);

return pos;

}

function bestPath(){

var best=nodes[0];

nodes.forEach(function(n){ if(n.score>best.score) best=n; });

var set={}, cur=best;

while(cur){ set[cur.id]=true; cur = cur.parent===null?null:nodes.find(function(x){return x.id===cur.parent;}); }

return set;

}

function render(){

var pos=layout();

var maxX=0,maxY=0;

for(var k in pos){ if(pos[k].x>maxX)maxX=pos[k].x; if(pos[k].y>maxY)maxY=pos[k].y; }

var padX=40, padY=34, gapX=Math.max(54, 520/(maxX+1)), gapY=64;

var W=padX*2 + maxX*gapX, H=padY*2 + maxY*gapY;

W=Math.max(W,560); H=Math.max(H,240);

svg.setAttribute("viewBox","0 0 "+W+" "+H);

while(svg.firstChild) svg.removeChild(svg.firstChild);

var best=bestPath();

var XY=function(id){return {x:padX+pos[id].x*gapX, y:padY+pos[id].y*gapY};};

// edges

nodes.forEach(function(n){

if(n.parent===null) return;

var a=XY(n.parent), b=XY(n.id);

var e=document.createElementNS(SVGNS,"path");

e.setAttribute("d","M"+a.x+","+a.y+" C"+a.x+","+(a.y+gapY/2)+" "+b.x+","+(b.y-gapY/2)+" "+b.x+","+b.y);

e.setAttribute("class","mtp-edge"+((best[n.id]&&best[n.parent])?" best":""));

svg.appendChild(e);

});

// nodes

nodes.forEach(function(n){

var p=XY(n.id);

var g=document.createElementNS(SVGNS,"circle");

g.setAttribute("cx",p.x); g.setAttribute("cy",p.y); g.setAttribute("r",n.id===0?13:11);

// fill = score ramp toward NVIDIA green

var t=n.score; var r=Math.round(42+(118-42)*t), gg=Math.round(42+(185-42)*t), bb=Math.round(42+(0-42)*t);

g.setAttribute("fill","rgb("+r+","+gg+","+bb+")");

var stroke = $("mtp-multi").checked ? MODELS[n.model].c : (best[n.id]?"#76B900":"#555");

g.setAttribute("stroke", best[n.id]?"#76B900":stroke);

g.setAttribute("stroke-width", best[n.id]?3:1.5);

svg.appendChild(g);

var tx=document.createElementNS(SVGNS,"text");

tx.setAttribute("x",p.x); tx.setAttribute("y",p.y+3.5); tx.setAttribute("text-anchor","middle");

tx.setAttribute("font-size","9"); tx.setAttribute("fill", t>0.55?"#0a0a0a":"#e8e8e8");

tx.textContent = n.id===0?"root":n.score.toFixed(2);

svg.appendChild(tx);

});

// stats

var bestScore=0; nodes.forEach(function(n){if(n.score>bestScore)bestScore=n.score;});

$("mtp-budget").textContent=used+" / "+BUDGET;

$("mtp-nodes").textContent=nodes.length;

$("mtp-best").textContent=bestScore.toFixed(2);

$("mtp-wd").textContent=widers+" / "+deepers;

$("mtp-step").disabled = used>=BUDGET;

if(used>=BUDGET && timer){clearInterval(timer);timer=null;$("mtp-run").textContent="✓ Done";}

}

$("mtp-run").onclick=function(){

if(used>=BUDGET) reset();

if(timer){clearInterval(timer);timer=null;this.textContent="► Run";return;}

this.textContent="⏸ Pause";

var self=this;

timer=setInterval(function(){ if(!step()){clearInterval(timer);timer=null;self.textContent="✓ Done";} },520);

};

$("mtp-step").onclick=function(){ if(used

How Marlin Compares

Marlin competes on depth, not speed. Conventional deep-research tools answer in minutes to tens of minutes. Marlin deliberately spends hours to raise output quality. The competitor run times below are approximate and reported, not official figures.

ToolTypical run timeOutputPrimary user

Sakana MarlinUp to ~8 hoursReport (dozens to ~100 pages) + slidesEnterprise strategy teams

OpenAI Deep Research~Minutes to tens of minutesCited text reportGeneral and pro users

Perplexity Deep Research~A few minutesCited text answerGeneral users

Google Gemini Deep Research~MinutesCited text reportGeneral and workspace users

The trade-off is explicit. You wait longer and pay per run. In return you get deeper hypothesis testing and a finished deliverable. You can cancel a run anytime, but credits are still consumed.

Pricing

Sakana offers pay-as-you-go along with Pro, Team, and Enterprise tiers. Pay-as-you-go starts at 100 credits per run, at ¥98 per credit. Pro is ¥150,000 per month and includes 2,000 credits. Team is ¥400,000 per month and includes 6,000 credits. Enterprise pricing is custom, with dedicated support.

Use Cases, With Examples

Marlin suits high-stakes questions where research is the bottleneck. Here are concrete examples drawn from its target tasks.

Market entry: 'Assess Japan's stablecoin and tokenized-payments market after regulatory change.' Marlin maps drivers, risks, and structured options into a report.

Risk analysis: 'Model resolution scenarios for a Strait of Hormuz blockade.' It compares hypotheses, not just summaries, before drawing conclusions.

Competitive analysis: Profile three rivals and rank our positioning gaps. It returns slides ready for a strategy review.

Each example fits one prompt and one unattended run. A human still reviews the cited output before any decision.

Try the Engine Yourself: TreeQuest

You cannot self-host Marlin. But you can run its core algorithm today. Sakana open-sourced AB-MCTS as TreeQuest under the Apache 2.0 license. Install it, define a generate function, then run a fixed search budget.

Copy CodeCopiedUse a different Browser

import random

import treequest as tq

Each node holds a user-defined state; score must be normalized to [0, 1].

def generate(parent_state):

if parent_state is None: # None means expand from the root

new_state = "Initial draft"

else:

new_state = f"Refined: {parent_state}"

score = random.random() # swap this for an LLM-based score

return new_state, score

algo = tq.ABMCTSA() # Adaptive Branching MCTS (variant A)

search_tree = algo.init_tree()

for _ in range(10): # generation budget of 10

search_tree = algo.step(search_tree, {"generate": generate})

best_state, best_score = tq.top_k(search_tree, algo, k=1)[0]

print("BEST:", best_state, round(best_score, 3))

Swap the random score for an LLM judge to reproduce the real pattern. TreeQuest also ships multi-LLM search and checkpointing for long runs. Checkpointing matters because long sessions can hit API errors midway.

Strengths and Weaknesses

Strengths

Peer-reviewed foundations: AB-MCTS at NeurIPS and AI Scientist in Nature.

Finished deliverables, including references, appendices, and slides.

Adaptive compute spends effort on the most promising branches.

The open-source core (TreeQuest) lets AI researchers study the method.

Weaknesses

Long runtimes make iteration slow versus minute-scale research tools.

Automated reports can contain hard-to-spot errors that need human review.

Pricing and design target enterprises, not individual developers.

Marlin itself is closed; only the underlying algorithm is open.

Key Takeaways

Sakana Marlin runs autonomous research for up to about eight hours per task.

One run produces a report of dozens of pages, plus slides.

It builds on AB-MCTS (NeurIPS 2025 Spotlight) and AI Scientist workflows (Nature).

Entry pricing is pay-as-you-go: 100 credits per run at ¥98 per credit.

It targets finance, corporate strategy, consulting, and think-tank teams.

Sources

Sakana AI — Sakana Marlin release: https://sakana.ai/marlin-release/

Sakana AI — Sakana Marlin product page: https://sakana.ai/marlin/

Sakana AI — AB-MCTS research and TreeQuest: https://sakana.ai/ab-mcts/

SakanaAI/treequest (GitHub, Apache 2.0): https://github.com/SakanaAI/treequest

The post Sakana AI Commercializes AB-MCTS in Sakana Marlin, an Enterprise Agent Generating Up to 100-Page Research Reports With Slides appeared first on MarkTechPost.