Sakana AI Opens Applied Research Lab for Market and Economic Modeling
Sakana AI has opened an Applied Research Lab inside the company, a group that applies its research to complex systems such as markets and economies. A single LLM agent cannot explain a market: it is modeled as a network of participants and links, while humans still set the questions and judge the answers. The first program covers finance and economics, and the team is being recruited in Tokyo. For agent builders, this signals a shift toward simulations built from many agents.
AI-processed from Sakana AI Blog; edited by Hamidun News
Japan's Sakana AI has announced the creation of an Applied Research Lab, a unit that will take the company's research and apply it to large, complex systems. The first target is finance and economics. The news is interesting not for its headline numbers, of which the announcement has none, but for how the company frames the task: a market cannot be "solved" with one clever chatbot.
Why one agent cannot understand a market
Sakana AI starts from a simple idea. A market or an economy is a system: it has individual participants, links between them, and chain reactions in which a change in one place echoes in another. To describe such a system, asking a language model a question is not enough. You need a model that shows who interacts with whom and what happens if something changes.
In the past, human experts chose the questions: what to model and how to check the result. That is why progress moved at the speed of human work. Today agentic AI can mass-produce, test and refine models, but it is weaker at the central task of deciding which questions matter. The lab aims to combine these two halves: human knowledge as the foundation and AI as the accelerator.
The first program is called Computational Finance and Economics, or COFE. Finance traditionally studies asset prices, while economics studies the system beneath them. According to the authors, AI makes it possible to connect both levels at the scale of individual companies and deals, almost in real time, so that prices become a visible reflection of the real economy.
What the lab builds on
The lab rests on work Sakana AI has done since its founding in July 2023: automating scientific discovery, evolutionary search, agent-based simulation and model evaluation. Its closest predecessors are the company's finance projects:
- EDINET-Bench tests LLMs on ten years of annual reports from Japanese companies. The unexpected result: even strong models only slightly outperformed logistic regression at detecting fraud and forecasting profits. The caveat is that the test is single-step and does not involve agents.
- JFinBench consists of exams for 2026 Japanese financial certifications, including audit and tax.
- CoffeeBench, built together with KPMG AZSA, examines whether an agent can run a simulated coffee supply chain profitably for 90 days.
- QD-Rec applies "quality plus diversity" search to decisions such as mergers and acquisitions: instead of a single ranking, it produces a set of different, justified options.
One of the first directions is agent-based modeling. The Shachi framework runs simulations in which each LLM agent is assembled from settings, memory and tools. Such a testbed can be used to study how collective behavior emerges from the actions of individual participants, for example how a market reacts to the imposition of tariffs. CEDAR adds LLM-guided tree search: it searches for the structure of a complex system so that it behaves as desired. For the algorithms themselves, Sakana AI uses ShinkaEvolve and AB-MCTS, which evolve code such as a trading strategy. Alongside them is The AI Scientist, which automates the path from hypothesis to paper.
What this means in practice
For agent developers, the takeaway is that value is shifting from single assistants to multi-agent systems with a shared environment, memory and metrics. If you build simulations, fintech analytics or decision-support tools, it is worth looking at Sakana AI's open work: Shachi and CEDAR are available on arXiv.
There are also honest caveats. The announcement is an organizational statement, not a product release: it contains no lab results, timelines or terms for customers. EDINET-Bench itself showed that on financial tasks, LLMs have so far barely moved beyond simple statistics, so the promise of "understanding the economy" still has to be demonstrated. The authors also point to reflexivity: markets react to those who watch them, which complicates any forecast.
The lab is located at Sakana AI's headquarters in Tokyo and is assembling a small interdisciplinary team. It is looking for researchers in physics, statistics, mathematics or economics with strong Python skills and experience with real-world data, as well as engineers who build data pipelines, simulation infrastructure and agent orchestration. Candidates can join through a Member of Technical Staff internship or the Visiting Researcher program. Sakana AI also invites collaboration with institutions, investors, regulators and companies that have complex practical problems.
- Nov 21, 2026Skein Agent Erases Finished Stages From Context: 97% Success Rate, But More Expensive Than OpenCode
- Nov 21, 2026Instead of Cursor: Six Open-Source Coding Agents, OpenCode Has 228K Stars on GitHub
- Nov 21, 2026ClawDOS Writes and Runs Code on an IBM PC 5160: Primes Under 100 in 64 Seconds
- Nov 21, 2026Apple Is Building an AI Agent and Making It Harder for Outsiders to Get Access: Permission Now Takes 14 Steps, Not 4
Key Questions Answered
Why one agent cannot understand a market?
Sakana AI starts from a simple idea. A market or an economy is a system: it has individual participants, links between them, and chain reactions in which a change in one place echoes in another. To describe such a system, asking a language model a question is not enough. You need a model that shows who interacts with whom and what happens if something changes. In the past, human experts chose the questions: what to model and how to check the result. That is why progress moved at the speed of human work.
What the lab builds on?
The lab rests on work Sakana AI has done since its founding in July 2023: automating scientific discovery, evolutionary search, agent-based simulation and model evaluation. Its closest predecessors are the company's finance projects: EDINET-Bench tests LLMs on ten years of annual reports from Japanese companies. The unexpected result: even strong models only slightly outperformed logistic regression at detecting fraud and forecasting profits. The caveat is that the test is single-step and does not involve agents.
What this means in practice?
For agent developers, the takeaway is that value is shifting from single assistants to multi-agent systems with a shared environment, memory and metrics. If you build simulations, fintech analytics or decision-support tools, it is worth looking at Sakana AI's open work: Shachi and CEDAR are available on arXiv. There are also honest caveats. The announcement is an organizational statement, not a product release: it contains no lab results, timelines or terms for customers.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.