Sakana Fugu - A multi-agent orchestration system launched by Sakana AI

What is Sakana Fugu?

Sakana Fugu is a multi-agent orchestration system from Sakana AI that dynamically schedules top-tier models using a single API. The system assigns thinker, executor, and verifier roles, automatically completing selection, delegation, and synthesis without requiring pre-defined workflows. Sakana Fugu supports recursive calls and model pool management, offering standard and Ultra versions. It achieves cutting-edge performance on stringent benchmarks, comparable to Fable 5 and Mythos Preview, without relying on a single vendor or facing export control risks.

Main functions of Sakana Fugu

• Unified scheduling via a single API: Users can invoke the OpenAI-compatible endpoint, and Fugu automatically completes model selection, task delegation, verification, and answer synthesis internally, without the need for a proprietary SDK.
• Dynamic role allocation: The system learns autonomously to assign the roles of thinker, executor, and verifier to different models, without the need for manual workflow pre-setting.
• Recursive self-calling: It supports calling its own instance as a coordinator to handle complex tasks that require multiple rounds of deep collaboration.
• Flexible model pool management: Users can choose to include or exclude specific vendor models to meet data privacy and compliance requirements.
• Dual version coverage scenarios: Fugu Standard balances performance and latency, while Fugu Ultra coordinates a deeper pool of experts to maximize the quality of answers for challenging tasks.
• Resilient and risk-resistant architecture: The underlying model pool is replaceable, and automatic routing bypass is performed when a single model is unavailable, avoiding service interruption and exit control risks.

Technical principles of Sakana Fugu

• TRINITY Evolutionary CoordinatorBased on the ICLR 2026 paper "TRINITY", the lightweight evolution module manages multiple LLMs in multiple rounds, adaptively assigning thinker, worker, and verifier roles and dynamically delegating tasks according to task types such as coding, mathematics, and reasoning, thus avoiding the efficiency bottleneck of fixed pipelines.
• Conductor reinforcement learning orchestrationBased on the ICLR 2026 paper "Conductor", we trained the coordinator to autonomously discover collaborative strategies at the natural language level through reinforcement learning, designed inter-agent communication patterns and focused prompts, and enabled a diverse cluster of models to outperform any single worker on difficult inference benchmarks.
• Model, i.e., orchestrator architectureFugu itself is a specially trained language model that understands when to delegate, how to communicate, and how to integrate. It predicts the next model and role to be invoked from the hidden state through a lightweight head, achieving efficient routing and collective intelligence emergence with extremely low additional computational cost.

How to use Sakana Fugu