Sakana Fugu: Multi-Agent System as a Model

Short summary

Sakana AI's Fugu system uses a multi-agent architecture that appears as a unified model from the outside while internally coordinating multiple specialist agents. A single API call triggers direct answering, specialist delegation, intermediate verification, and final synthesis. This represents a shift away from traditional scaling approaches focused solely on larger parameters and extended context windows.

• •Fugu presents as a single model but runs multiple coordinated agent specialists internally
• •Unified API enables direct answering, delegation, verification, and synthesis in one workflow
• •Diverges from traditional LLM scaling focused on parameters and context length

Generated with AI, which can make mistakes.