Sakana Fugu - A Multi-Agent Orchestration System Launched by Sakana AI

What Is Sakana Fugu

Sakana Fugu is a multi-agent orchestration system launched by Sakana AI that dynamically dispatches top-tier models through a single API. The system assigns Thinker, Worker, and Verifier roles, automatically handling selection, delegation, and synthesis without preset workflows. Sakana Fugu supports recursive invocation and model pool governance, and offers a Standard edition and an Ultra edition, achieving frontier-level performance on demanding benchmarks, rivalingFable 5andMythos Preview, with no single-vendor dependency or export control risk.

Sakana Fugu's Main Features

• Unified dispatch via a single API: Users call through an OpenAI-compatible endpoint, and Fugu internally completes model selection, task delegation, verification, and answer synthesis, with no proprietary SDK required.
• Dynamic role assignment: The system autonomously learns to assign Thinker, Worker, and Verifier roles to different models, without manually preset workflows.
• Recursive self-invocation: Supports invoking its own instances as coordinators to handle complex tasks requiring multiple rounds of deep collaboration.
• Flexible model pool governance: Users can independently choose to include or exclude specific vendors' models to meet data privacy and compliance requirements.
• Two editions covering scenarios: Fugu Standard balances performance and latency, while Fugu Ultra coordinates a deeper expert pool to maximize answer quality on high-difficulty tasks.
• Resilient risk-resistant architecture: The underlying model pool is replaceable, and when a single model is unavailable, routing automatically bypasses it, avoiding service interruptions and export control risks.

Sakana Fugu's Technical Principles

• TRINITY evolutionary coordinator: Based on the ICLR 2026 paper "TRINITY", a lightweight evolutionary module governs multiple LLMs across multiple rounds, adaptively assigning Thinker, Worker, and Verifier roles according to task types such as coding, mathematics, and reasoning, and dynamically delegating work to avoid the efficiency bottlenecks of fixed pipelines.
• Conductor reinforcement learning orchestration: Based on the ICLR 2026 paper "Conductor", a coordinator is trained through reinforcement learning to autonomously discover collaboration strategies at the natural language level, designing inter-agent communication patterns and focused prompts, enabling diverse model clusters to surpass any single worker on difficult reasoning benchmarks.
• Model-as-orchestrator architecture: Fugu itself is a specially trained language model that understands when to delegate, how to communicate, and how to integrate; through a lightweight head, it predicts from hidden states which model and role should be invoked next, achieving efficient routing and the emergence of collective intelligence at extremely low additional computational cost.

How to Use Sakana Fugu