Sakana Fugu is a sophisticated multi-agent system presented as a unified LLM, designed to tackle complex, multi-step tasks by dynamically orchestrating a diverse pool of frontier models. It offers access via the Sakana API, supporting Chat Completions and Responses endpoints, and provides a significant benefit by achieving superior performance through intelligent coordination. This system is targeted at developers and researchers seeking advanced LLM capabilities for intricate problem-solving.

How It Works

Sakana Fugu operates as a multi-agent system integrated into a single LLM interface. Its core innovation lies in dynamic orchestration, leveraging two key research components: TRINITY, an evolutionary-optimized coordinator that delegates tasks turn-by-turn to a pool of LLMs without weight merging, and the Conductor, an RL-trained model that designs agent communication topologies and crafts targeted instructions. This approach enables sophisticated collaboration among diverse frontier models, discovering coordination strategies that surpass individual model capabilities.

Quick Start & Requirements

• Primary install / run command:
  • Install: curl -fsSL https://sakana.ai/fugu/install | bash
  • Launch: codex-fugu
• Non-default prerequisites and dependencies: Requires the Codex environment. The one-line install supports Ubuntu and macOS.
• Links: Installation script (https://sakana.ai/fugu/install), Support (https://fugu.sakana.ai).

Highlighted Details

• Achieves superior performance on complex, multi-step tasks through intelligent coordination of frontier models.
• Continuously updated model pool and retrained coordinators ensure performance advantage as new frontier models emerge.
• Evaluation details and performance claims are documented in a technical report (arXiv:2606.21228), reflecting June 2026 benchmarks.
• Based on foundational research from two ICLR 2026 papers: "TRINITY: An Evolved LLM Coordinator" and "Learning to Orchestrate Agents in Natural Language with the Conductor."

Maintenance & Community

• Support contact for issues or bugs: https://fugu.sakana.ai.

Licensing & Compatibility

• No explicit license information is provided in the README.
• Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The one-line installation script is explicitly supported for Ubuntu and macOS; users on Windows or those encountering installation issues must refer to a separate guide, implying potential setup complexities or platform-specific configurations. No explicit mention of licensing terms or compatibility for commercial use is provided in the README, requiring further investigation.