Sakana Marlin: the autonomous “Virtual CSO” that runs unattended research and hands you a finished report

• What it is: Sakana Marlin is an enterprise research agent — a “virtual CSO” — that runs unattended sessions (up to about eight hours) and returns long, evidence‑backed reports and slide decks for finance, strategy, consulting and policy teams.
• How it works, simply: Marlin uses AB‑MCTS (Adaptive Branching Monte Carlo Tree Search) to balance exploring many hypotheses and drilling deeply into promising ones; it can also route steps to different LLMs to play to each model’s strengths.
• Main benefits: compresses weeks of desk research into a single unattended run, produces citation‑rich deliverables (dozens to ~100 pages plus slides), and checkpoints long searches so work can resume.
• Key risks: long runtimes slow iteration, automated outputs can include subtle errors that need verification, and the product is closed commercial software that may create vendor lock‑in and governance questions.

Quick hook

Replace a three‑week strategy sprint with a single overnight run that returns a fully sourced dossier and a ready slide deck — that’s the value proposition Marlin sells. It’s not a chat assistant; it’s an automated research lab that works while your team sleeps.

What Sakana Marlin does and who should care

Sakana Marlin is an autonomous enterprise research agent targeted at corporate strategy, investment teams, consultancies and think tanks. A typical run lasts up to about eight hours and issues hundreds to thousands of LLM queries. The output is a finished, citation‑backed report (dozens to roughly 100 pages), appendices, and a presentation deck with images generated by image‑capable models. Beta testers — roughly 300 professionals in April 2026 — used Marlin for strategy, risk, competitive intelligence and market research.

“Marlin is positioned as a Virtual CSO — an autonomous strategic research engine for enterprises.”

How it works: AB‑MCTS and multi‑LLM routing (plain English)

AB‑MCTS (Adaptive Branching Monte Carlo Tree Search) reframes reasoning as a branching search. Think of it as a research team that splits into multiple squads: some squads explore new leads (go wider), while others dig deep on the most promising evidence (go deeper). The system dynamically reallocates compute the way a manager reallocates headcount to winning investigations.

Key technical points without the jargon: AB‑MCTS balances exploration vs. exploitation at inference time rather than relying on many blind samples. Marlin can also route specific tasks to different LLMs (multi‑LLM routing) — sending summary work to a cheaper, faster model and complex synthesis or nuance checks to a larger specialist model. This mix improves task completion metrics in Sakana’s experiments: combining models solved about 27.5% of benchmark tasks versus about 23% for one small model alone — a measurable uplift for difficult, long‑horizon problems.

“AB‑MCTS reframes reasoning as a tree search that can choose to go wider or deeper.”

Rough runtime flow

1. Seed prompt / brief from the user.
2. Branching stage: spawn exploratory threads for hypotheses and data sources.
3. Checkpointing: save progress so long runs (hours) can be resumed.
4. Routing: send sub‑tasks to different LLMs as appropriate.
5. Synthesis and citation: assemble findings into a narrative, attach sources and appendices.
6. Generate slides and images for presentation output.

What a Marlin run returns

Expect a finished product rather than a conversation transcript. Typical deliverables include:

• Executive summary and strategic recommendations.
• Main report (dozens to ~100 pages) with evidence, quotes, tables and appended sources — hands‑on runs cited 60–80 sources.
• Slide deck (editable PPTX or similar) with image generation for visualizations.
• Appendices and raw search logs/checkpoints for auditability and review.

Sample table of contents for a Marlin report:

• Executive summary
• Market overview and TAM estimates
• Competitor landscape and positioning
• Financial implications and scenario models
• Operational risks and mitigations
• Sources and appendices (links to all cited documents)

Pricing (what procurement needs to know)

• Pay‑as‑you‑go: entry runs start at 100 credits per run, priced at ¥98/credit. A 100‑credit run costs ¥9,800 (≈USD 65, approx.).
• Pro: ¥150,000/month for 2,000 credits (≈USD 1,000/month, approx.).
• Team: ¥400,000/month for 6,000 credits (≈USD 2,667/month, approx.).
• Enterprise: custom pricing with SLAs and integrations.

Note: a single run can consume many credits depending on duration and model mix. Factor in credits plus human review hours when sizing pilots.

When to use Marlin — and when not to

When to use it

• Market entry and TAM analysis that require deep source collection and synthesis.
• M&A pre‑screening or due diligence where a fast, evidence‑rich first draft accelerates decisions.
• Geopolitical or scenario risk modeling that benefits from branching exploration of outcomes.
• Competitive intelligence and strategic positioning where a finished deck speeds board conversations.

When not to use it

• Rapid hypothesis testing or exploratory brainstorming — chat assistants and short analytic cycles are faster and cheaper.
• Tasks requiring near‑real‑time iteration or constant back‑and‑forth with stakeholders.
• Cases where you cannot tolerate any automated errors without exhaustive human‑in‑the‑loop checks (early pilots recommended).

Governance, accuracy and operational risks

Marlin brings automation savings but also governance responsibilities. Enterprise buyers should verify:

• Provenance & auditability: Are all claims traceable to sources and are links preserved in appendices? Can the system export provenance logs?
• Factual accuracy sampling: How does the product surface uncertainty or conflicting sources? Does it flag low‑confidence claims?
• Data handling: Where is sensitive input processed and stored? Are there options for on‑prem or VPC deployments for regulated sectors?
• Model dependencies: Which third‑party LLMs are used, and how will changes in model access or pricing affect total cost?
• Explainability and audit trails: Are the search checkpoints and the branching decisions logged so reviewers can follow the reasoning path?
• Regulatory compliance: GDPR, financial regulations, and sector rules — confirm data residency, retention and deletion policies.

Pilot evaluation checklist (what to test in a POC)

1. Factual accuracy sampling: Randomly verify 20–30 claims against ground truth; acceptance criteria: ≥95% traceable and ≤5% substantial factual error.
2. Source provenance test: All cited claims must link to a verifiable source or document; acceptance: 100% linkability for top‑level claims.
3. Revision latency: Measure time to iterate on a report after reviewer feedback; acceptance: reasonable turnaround (target depends on use case).
4. Cost & credits burn: Run representative tasks and measure credits per run; acceptance: predictable credit usage within budget.
5. Integration export: Ensure PPTX, PDF and raw logs export cleanly to your workflows (CMS, DAM, BI tools).
6. Security posture: Pen test / data handling review if PII or regulated data is used.

Competitive context and strategic implications

Marlin represents a class of AI agents that trade conversational speed for long‑horizon depth. It sits between quick chat assistants (ChatGPT‑style) and human analyst teams. Compared with search/aggregation tools (e.g., market intelligence platforms), Marlin emphasizes narrative synthesis and finished deliverables rather than surfacing snippets and alerts.

Strategically, this genre will shift how teams buy research: procurement will increasingly ask for pilot metrics (accuracy, credits per run, SLA) rather than seats alone. Internally, skill requirements shift from pure data gathering to verification, interpretation and strategic judgment — humans remain essential oversight.

Bottom line for executives

Autonomous research agents like Marlin show the next phase of AI for business: unattended, long‑horizon execution that hands you an actionable product. They can accelerate decision cycles and free analysts from repetitive collection work, but they demand a strong verification and governance posture. Treat the first pilots as a verification exercise focused on accuracy, provenance and total cost of ownership.

FAQ

What is an autonomous research agent?

An autonomous research agent is an AI system designed to run multi‑step research workflows unattended, producing final deliverables like reports and slide decks instead of chat responses. It combines search, reasoning and synthesis over long runs.

How does AB‑MCTS differ from ChatGPT‑style assistants?

Chat assistants typically answer in short, interactive turns. AB‑MCTS treats reasoning as a branching search — it explores multiple hypotheses and funnels compute into the best leads over a long run, which produces deeper, more evidence‑backed outputs at the cost of iteration speed.

Is any part open source?

The AB‑MCTS core is available as TreeQuest under an open‑source license. The Marlin commercial product, with integrations, UX and enterprise features, remains closed.

How reliable are the reports?

Marlin can produce citation‑rich reports, but automated synthesis can still produce subtle errors. Plan for human verification to sample claims, check provenance and validate key conclusions.

Who should evaluate Marlin first?

Strategy teams, M&A groups, competitive intelligence units and consultancies that need deep, evidence‑backed deliverables and can budget for both compute/credits and the human review loop.

Next practical steps

• Run a focused pilot: pick a representative, high‑value topic and test factual accuracy, provenance and export flows.
• Budget for human review: include verification hours in total cost of ownership calculations.
• Validate governance: get clear answers on data residency, model dependencies and audit logs before production use.

Autonomous research agents won’t replace judgment, but they will change how research is produced. The winning teams will be those that pair these tools with rigorous verification processes and clear decision‑rights — letting AI shoulder the grunt work while humans raise the strategic questions.