Sakana AI Launches 'Marlin', an Autonomous Agent That Spends 8 Hours Writing 100-Page Strategy Reports

For the past two years, the generative AI industry has been locked in a race for speed, measuring success by how many milliseconds it takes a chatbot to spit out a poem or a code snippet. On Monday, Tokyo-based startup Sakana AI deliberately ran in the opposite direction.[1]

The company officially launched "Sakana Marlin," its first commercial product, which it bills as a "Virtual Chief Strategy Officer" (CSO). Instead of delivering instant, surface-level answers, Marlin is an autonomous B2B research agent designed to run continuous, self-governing reasoning loops for up to eight hours at a time.[1][5]

The output of this extended computational marathon is not a quick chat bubble, but a comprehensive, heavily cited strategy report that can stretch to roughly 100 pages, accompanied by executive summary slides.[1][2]

"With Marlin, major businesses are no longer asking how fast an AI can answer, but how deeply it can think," noted industry observers, highlighting a paradigm shift from shallow generation to methodical reasoning.[1]

The workflow is fundamentally different from typical large language model interactions. Rather than engaging in a tedious back-and-forth prompt engineering session, a user simply provides a core research topic and steps away.[1][5]

For the next several hours, Marlin operates as a self-contained digital strategy team. It formulates its own initial hypotheses, navigates the web to gather data, cross-references conflicting information, and maps out causal relationships in complex business environments.[1][5]

The technology underpinning Marlin grew out of two notable research efforts at Sakana AI. The first is the "AI Scientist" framework, an automated research system that earned a feature in the journal Nature.[2][5]

The second is AB-MCTS (Adaptive Branching Monte Carlo Tree Search), a reasoning method spotlighted at the NeurIPS 2025 conference. This technique treats the reasoning process as a "tree search," allowing the AI to autonomously evaluate the validity of various hypotheses and concentrate its computational resources on the most promising paths.[2][6]

By scaling up "inference-time compute"—the processing power used while the model is actually generating an answer, rather than during its initial training—Marlin effectively thinks longer to produce better results.[5][7]

Before its public launch, Sakana AI ran a closed beta starting in April 2026 with approximately 300 professionals from sectors including finance, consulting, and think tanks. The goal was to test the agent on real-world tasks like geopolitical risk analysis, market research, and supply chain vulnerability assessments.[2][6]

The commercial release targets the enterprise market with a dual pricing structure: a pay-as-you-go model costing 100 credits per run, and monthly subscription plans starting at ¥150,000 (roughly $1,000).[2][6]

Sakana AI, founded in 2023 by former Google researchers, is well-capitalized to push this enterprise frontier. Following a $135 million Series B funding round in late 2025, the company reached a post-money valuation of approximately $2.65 billion, backed by Silicon Valley's Khosla Ventures and Japan's largest financial group, MUFG.[2][7]

Despite the impressive capabilities, analysts caution that autonomous agents running for hours introduce new categories of risk. The primary concern is "hallucinations at scale."[3][7]

If a standard chatbot fabricates a fact in a two-paragraph response, it is easily spotted and corrected. But if an autonomous agent establishes a flawed premise on page 12 of a 100-page report, it could spend the next seven hours and 88 pages compounding that error into a highly convincing, yet fundamentally wrong, strategic recommendation.[3][7]

To mitigate these risks, Sakana AI emphasizes that Marlin is designed to handle the heavy lifting of comprehensive research, but the final, highest-value work—the actual decision-making—must remain in human hands.[5][8]