Sakana AI Launches Fugu: Multi-Agent AI System Built on Collective Intelligence Research and Frontier Model Coordination

Sakana AI has introduced Sakana Fugu, a new AI product built around multi-agent orchestration. The system is designed to coordinate multiple frontier foundation models in order to deliver strong results across coding, mathematics, scientific reasoning, and other demanding tasks.

The product is initially being released through an API, reflecting its use as an internal tool for Sakana AI’s own researchers and engineers before becoming available to external users. The company is positioning the launch as a step toward a broader commercial framework for systems that do not rely on a single large model, but instead combine specialized models and agents that work together.

Sakana AI has framed Fugu as part of a wider research direction centered on collective intelligence. That approach includes earlier projects such as evolutionary model merging, which explored how diverse open-source models could be combined to create capabilities not found in any single system, as well as The AI Scientist, a system that demonstrated coordinated agents carrying out the full research cycle autonomously. 

The company has also pointed to ShinkaEvolve, which uses evolutionary search over LLM-generated programs to identify algorithms that outperform human-written solutions, and AB-MCTS, which showed that multiple frontier models working together through tree search could outperform individual models on difficult reasoning tasks. Sakana Fugu is presented as the commercial expression of that body of work.

Multi-Agent Orchestration, Research Foundations, And API Compatibility In Fugu 

According to Sakana AI, the product is intended to reduce the complexity of working with multiple model providers. In many current workflows, users must manage several API keys and switch between models that each perform best in different areas. Fugu addresses this by dynamically coordinating a pool of models, rather than requiring users to manually define roles, workflows, or model assignments. The system is designed to learn how to assemble agents and distribute tasks in ways that improve efficiency and overall performance.

The models are based on Sakana AI’s ICLR 2026 research papers, titled Trinity and Conductor, with further refinements made for the commercial release. The company says the beta results already indicate competitive benchmark performance, and that the product is intended to support both latency-sensitive and performance-focused use cases through two variants: Fugu Mini, optimized for speed, and Fugu Ultra, built for more demanding workloads.

Access is available through APIs compatible with standard OpenAI-style endpoints, allowing integration into existing workflows with limited changes. 

The post Sakana AI Launches Fugu: Multi-Agent AI System Built on Collective Intelligence Research and Frontier Model Coordination appeared first on Metaverse Post.