Sakana Fugu multi-agent system delivered as one model (github.com)

Sakana has introduced Fugu, a cutting-edge multi-agent system integrated into a single model that dynamically orchestrates various frontier models to address complex, multi-step tasks. This innovative system can be accessed via the Sakana API, supporting interactions through both Chat Completions and Responses endpoints. Users can easily set up Fugu by executing a simple installation command for quick integration into Codex, enabling immediate engagement with its capabilities. The significance of Sakana Fugu for the AI/ML community lies in its superior performance achieved through the coordination of a diverse range of powerful models. This orchestration is facilitated by two advanced models: the TRINITY coordinator, which uses an evolutionary strategy to allocate roles among models, and the Conductor, trained via reinforcement learning to optimize agent communication and instruction flow. By continuously updating its model pool with the latest frontier releases and retraining its coordinators, Fugu maintains a competitive edge against other leading models, as evidenced by its favorable comparisons with unnamed frontiers in recent evaluations. The ongoing enhancements and detailed technical reports ensure that Fugu remains at the forefront of AI advancements.