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

Sakana AI has launched Sakana Fugu, a brand new AI product constructed round multi-agent orchestration. The system is designed to coordinate a number of frontier basis fashions with a view to ship sturdy outcomes throughout coding, arithmetic, scientific reasoning, and different demanding duties.

The product is initially being launched by means of an API, reflecting its use as an inside device for Sakana AI’s personal researchers and engineers earlier than turning into out there to exterior customers. The firm is positioning the launch as a step towards a broader business framework for methods that don’t rely on a single massive mannequin, however as a substitute mix specialised fashions and brokers that work collectively.

Sakana AI has framed Fugu as a part of a wider analysis route centered on collective intelligence. That method consists of earlier tasks reminiscent of evolutionary mannequin merging, which explored how various open-source fashions might be mixed to create capabilities not present in any single system, in addition to The AI Scientist, a system that demonstrated coordinated brokers finishing up the complete analysis cycle autonomously. 

The firm has additionally pointed to ShinkaEvolve, which makes use of evolutionary search over LLM-generated applications to establish algorithms that outperform human-written options, and AB-MCTS, which confirmed that a number of frontier fashions working collectively by means of tree search might outperform particular person fashions on troublesome reasoning duties. Sakana Fugu is offered because the business expression of that physique of labor.

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

According to Sakana AI, the product is meant to scale back the complexity of working with a number of mannequin suppliers. In many present workflows, customers should handle a number of API keys and change between fashions that every carry out finest in several areas. Fugu addresses this by dynamically coordinating a pool of fashions, slightly than requiring customers to manually outline roles, workflows, or mannequin assignments. The system is designed to discover ways to assemble brokers and distribute duties in ways in which enhance effectivity and total efficiency.

The fashions are based mostly on Sakana AI’s ICLR 2026 analysis papers, titled Trinity and Conductor, with additional refinements made for the business launch. The firm says the beta outcomes already point out aggressive benchmark efficiency, and that the product is meant to help each latency-sensitive and performance-focused use instances by means of two variants: Fugu Mini, optimized for pace, and Fugu Ultra, constructed for extra demanding workloads.

Access is offered by means of APIs appropriate with customary OpenAI-style endpoints, permitting integration into present workflows with restricted modifications. 

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