The Shift to Long-Horizon AI: How Sakana Marlin Compresses Weeks of Strategy Research into Hours

On June 15, 2026, Tokyo-based AI startup Sakana AI officially launched its first commercial product, an autonomous B2B research agent named Sakana Marlin. Billed as a "Virtual Chief Strategy Officer" (CSO), the platform is designed to take on the heavy lifting of enterprise research, compressing what would typically take a human team weeks of work into a matter of hours.[1][5]

The generative AI hype cycle has largely been defined by speed, with the industry standard set by chatbots capable of generating surface-level summaries or code snippets in mere milliseconds. However, Sakana Marlin deliberately abandons this instantaneous text generation. Instead, it represents a paradigm shift toward deep, methodical analysis, proving that the enterprise frontier is moving from shallow generation to long-horizon reasoning.[1][5]

What sets Marlin apart from the current ecosystem of AI tools is its temporal scale. Rather than returning an answer immediately, the system runs continuous, self-governing reasoning loops for up to eight hours at a time. During this extended window, the AI operates as a self-contained digital strategy team, working without any further human intervention.[1][6]

The workflow is fundamentally different from typical prompt engineering. A user simply provides a core research topic and clarifies the direction through a brief initial conversation. From there, the human steps away entirely, allowing the agent to autonomously formulate its own initial hypotheses, navigate the web to gather data, and cross-reference sources.[1][4]

At the core of Marlin's capabilities is a technology known as Adaptive Branching Monte Carlo Tree Search (AB-MCTS). This algorithm allows the AI to treat the research process as a branching tree of possibilities, evaluating the validity of multiple hypotheses simultaneously. Rather than settling for the first plausible answer, Marlin continuously verifies its findings, pruning dead ends and concentrating its computational resources on the most promising analytical paths.[2][5]

This approach relies heavily on "inference-time scaling"—the concept of spending significantly more computational power during the generation phase rather than just during the model's initial training. By allowing the AI to "think" for hours, Sakana AI ensures that the final output is not just a regurgitation of secondary sources, but a structured synthesis of primary data and resolved contradictions.[5][6]

Marlin's architecture is also heavily informed by Sakana AI's previous breakthrough, the "AI Scientist" framework. Published in the journal Nature, that project demonstrated an AI capable of automating the entire scientific discovery cycle, from idea generation to peer review. Marlin adapts this autonomous workflow for the corporate world, applying the same rigorous, multi-model coordination to business strategy.[4][5]

The output generated by this exhaustive process is designed strictly for executive decision-makers. After an eight-hour run, Marlin delivers a deeply researched, fully cited strategy report that can span up to 100 pages. Alongside the comprehensive document, the system also generates an executive summary in the form of presentation slides, complete with structured strategic options.[1][2]

Prior to its official launch, Sakana AI refined Marlin through a rigorous closed beta testing phase that began in April 2026. Approximately 300 professionals across finance, corporate consulting, and think tanks tested the system on real-world tasks. Feedback from this cohort was instrumental in improving the agent's research depth, output formatting, and the stability of its long-duration reasoning tasks.[2][4]

During the beta, users deployed Marlin to tackle highly complex, multi-layered business challenges. Examples of its application included analyzing the geopolitical risks and supply chain impacts of a second Trump administration, assessing the regulatory implications of tokenized payments in Japan, and conducting deep competitive analyses of the enterprise AI market.[2][5]

To support these enterprise use cases, Sakana AI has structured Marlin as a strictly B2B service. The platform is available through tiered subscription plans, including Pro, Team, and Enterprise options, as well as a pay-per-use model. Recognizing the sensitive nature of corporate strategy, the company has also emphasized strict data privacy, ensuring that user inputs and proprietary research topics are not used to train their underlying models.[3][5]

The introduction of Marlin underscores a larger trend in the technology sector: the push for greater autonomy in business applications. By successfully commercializing long-horizon reasoning, Sakana AI is challenging the myth that a single, centralized "omnipotent" AI model is the only path forward. Instead, they are proving the viability of specialized, multi-agent ecosystems that divide labor to complete high-level tasks.[2][5]

For corporate strategists, the value proposition is clear. The goal is not to replace the human decision-maker, but to eliminate the research bottleneck that consumes weeks of human capital. By taking on the heavy lifting of exhaustive data gathering and structural analysis, Marlin frees human teams from the "gravity" of preliminary research.[3][6]

This reallocation of labor allows executives and strategists to concentrate on the highest-value work of all: the actual decision-making. When an AI can deliver a 100-page, review-quality report overnight, businesses can react to market shifts and complex challenges with unprecedented agility.[1][2]

As the enterprise AI market continues to mature, tools like Sakana Marlin represent a critical evolution. Major businesses are no longer asking how fast an AI can answer a question, but how deeply it can think about a problem.[1][2]

Ultimately, the success of this new paradigm will hinge on its real-world reliability. As more corporations integrate autonomous research agents into their strategic workflows, the industry will be watching closely to see if AI can consistently match the nuanced insights of human strategists while operating at machine speed.[2]