Sakana Fugu: Orchestration for Frontier LLMs

Sakana AI has launched Sakana Fugu, a new multi-agent orchestration system. This system acts like a single model, handling requests through one endpoint. Internally, Fugu decides whether to solve a task directly or coordinate a team of expert models. This means the complexity of a multi-agent system stays hidden from your code. Fugu is a language model itself, trained to call other large language models in an agent pool. It manages model selection, delegation, verification, and synthesis. Sakana AI states this approach helps prevent single-vendor dependency. There are two variants: Fugu, which balances performance and low latency for everyday tasks, and Fugu Ultra, designed for maximum quality on complex problems. Fugu Ultra leads most published coding and reasoning benchmarks. This innovation could simplify how developers interact with powerful AI systems, making advanced capabilities more accessible.