Sakana AI × Sumitomo Mitsui: Multi-AI Agent Collaboration Auto-Generates Proposals, Boosting Efficiency 100x

Sakana AI and SMBC Partnership: Proposal Creation Reduced from 1-2 Weeks to Hours

Sakana AI and Sumitomo Mitsui Financial Group (SMBC) have jointly developed a "Proposal Auto-Generation Application" that is now in active use at Sumitomo Mitsui Banking Corporation. This is the first deployed project since the two parties signed a partnership agreement in May 2025.

The most striking data point: proposal creation for large corporate clients, which previously took 1-2 weeks, can now be completed in tens of minutes to several hours. This isn't merely an efficiency improvement—it represents the deep application of AI Agents in complex business processes.

Sakana AI and SMBC jointly developed proposal auto-generation application

Technical Architecture Analysis: How Multi-AI Agent Autonomous Collaboration Works

Agent Clusters with Clear Role Division

The core technical highlight of this application is its multi-AI Agent autonomous collaboration architecture. Unlike simple Q&A with a single large model, multiple Agents in the system each handle specific responsibilities:

Information Gathering Agent: Automatically collects financial and non-financial information about target companies
Analysis Agent: Performs deep analysis on collected data
Hypothesis Building Agent: Proposes strategic hypotheses based on analysis results
Story Planning Agent: Organizes analysis and hypotheses into coherent proposal narratives
Quality Assessment & Fact-Checking Agent: Reviews and validates generated content

These Agents collaborate with each other, with AI autonomously constructing and executing optimal workflows, ultimately outputting highly consistent and professionally polished proposal content.

Technical Background of Multi-Agent Systems

Multi-Agent Systems (MAS) represent a core research direction in distributed artificial intelligence, with theoretical foundations tracing back to distributed problem-solving research in the 1980s. Unlike the Prompt Engineering approach with a single large language model, multi-Agent architectures decompose complex tasks into multiple subtasks, each handled by Agents with different capabilities and roles. Each Agent possesses independent reasoning ability, memory systems, and tool-calling permissions, collaborating through predefined or dynamically generated communication protocols. Since 2024, with the maturation of open-source frameworks like AutoGen, CrewAI, and LangGraph, multi-Agent systems have rapidly transitioned from academic concepts to engineering implementations. Their core advantages include: task decomposition reduces the complexity of individual reasoning steps, role specialization improves output quality, and inter-Agent review mechanisms provide built-in quality assurance.

Beyond Document Generation: AI-Driven Strategic Thinking Support

Interestingly, this application is not positioned as a simple "automated document generation tool." Sakana AI emphasizes that it achieves AI strategic thinking support—including hypothesis building and multi-perspective analysis. The system can propose new perspectives and objective arguments that humans might easily overlook, helping bank employees focus on solving clients' fundamental challenges.

Financial AI Agent Deployment: Why Proposals Are the Ideal Use Case

Natural Fit for Knowledge-Intensive Tasks

Proposal creation in wholesale banking is a quintessential knowledge-intensive task. It requires:

Extensive information research (industry trends, corporate financial reports, competitive landscape)
Professional financial analysis capabilities
Strategic thinking and hypothesis validation
High-quality document organization and presentation

This is precisely the ideal application scenario for multi-Agent systems—high task complexity, involvement of multiple professional domains, and the need for multi-step coordination. By comparison, a single Agent or simple RAG system would struggle to handle such a complex end-to-end process.

RAG (Retrieval-Augmented Generation) is currently the most widely adopted technical paradigm in enterprise AI applications. Its core approach combines external knowledge bases with large language models, retrieving relevant document fragments to enhance response accuracy. However, RAG is fundamentally still a two-step "retrieve + generate" process, suitable for answering factual questions or summarizing existing documents. When tasks involve multi-step reasoning, hypothesis validation, cross-domain information integration, and creative analysis, RAG systems reveal their limitations. Multi-Agent systems, by introducing capabilities like Planning, Reflection, and Tool Use, can handle complex workflows requiring multiple iterations and multi-angle verification—precisely the capability level needed for end-to-end tasks like proposal creation.

Industry Specifics of Japanese Corporate Banking

Japan's mega banks play a unique role in corporate banking. Unlike Western investment banks that focus on deal-making, Japanese banks' corporate business emphasizes "accompaniment-style" (伴走型) service—long-term tracking of client companies' business conditions while proactively proposing comprehensive solutions including management improvement, M&A, and financing. This model requires bank employees to possess deep industry knowledge and strategic analysis capabilities, with proposal quality directly impacting the depth of client relationships. However, Japanese banking has long faced the dual pressures of talent shortages and low work efficiency. According to the Japanese Bankers Association, bank employees spend approximately 30-40% of their working hours on document preparation and information gathering. The introduction of AI Agent technology aims to liberate human resources from repetitive information processing work, enabling them to focus on higher-value client communication and strategic judgment.

Sakana AI's Differentiated Technical Approach

Founded by former Google researchers, Sakana AI has consistently featured "nature-inspired AI" and multi-Agent collaboration as its technical hallmarks. This deployment with Japan's top financial group demonstrates the viability of its technical approach in enterprise-grade scenarios.

Sakana AI was established in Tokyo in 2023, co-founded by former Google Brain researcher David Ha and Llion Jones (one of the co-authors of the Transformer paper "Attention Is All You Need"). The company name comes from the Japanese word for "fish" (さかな), symbolizing its core philosophy—drawing inspiration from collective intelligence in nature (such as the collaborative behavior of fish schools and ant colonies) to design AI systems. The company's technical specialties include: evolutionary algorithm-driven Model Merging, multi-Agent collaboration frameworks, and research into achieving large model-level capabilities through clusters of smaller models. In 2024, Sakana AI completed over $300 million in funding at a valuation exceeding $2 billion, making it one of the highest-valued AI startups in Japan. Its investors include tech giants such as Google, NVIDIA, and NTT.

From a commercial perspective, this also represents an important step in Sakana AI's transformation from a research-focused company to a product-oriented one. Choosing finance—a high-value, high-barrier industry—as the entry point both demonstrates technical prowess and helps build commercial moats.

Future Outlook: From Bank Proposals to Full Financial Scenario Coverage

Sakana AI states that the Sumitomo Mitsui Banking application is just the beginning, with plans to gradually expand AI Agent technology applications across other business areas within the SMBC Group.

From a broader perspective, this case serves as a demonstration for the entire financial industry. When AI Agents can compress weeks of work into hours, bank employees' roles will shift from "information gatherers and document creators" to "strategic advisors and relationship managers." This is not merely an efficiency revolution—it's a fundamental transformation of the financial services model.

Global Financial AI Agent Competitive Landscape

The financial industry is one of the most active sectors for AI Agent technology deployment. Globally, JPMorgan Chase has deployed an internal AI platform called LLM Suite, Goldman Sachs has developed an AI assistant for investment banking operations, and Morgan Stanley has partnered with OpenAI to build an AI knowledge system for wealth management. Domestically in Japan, Mitsubishi UFJ Financial Group is advancing generative AI applications with Microsoft, while Mizuho Financial Group has introduced AI assistance in credit review. The uniqueness of SMBC's collaboration with Sakana AI lies in the fact that it doesn't simply embed a general-purpose large model into existing processes, but instead adopts the more cutting-edge multi-Agent collaboration architecture, attempting to achieve full-chain automation from information gathering to strategic analysis. The success or failure of this technical choice may influence the future direction of financial AI development in Japan.

This project was also covered by Nikkei (Japan's leading business newspaper), indicating the Japanese financial industry's strong interest in AI Agent technology. In the global AI competition, Japanese companies are accelerating the integration of cutting-edge AI technology into core business processes.