Sakana Fugu

Sakana Fugu is a multi-agent AI system that acts as a single model. Instead of relying on one large language model, it dynamically coordinates a pool of expert models — assigning them roles like thinker, worker, or verifier — to tackle complex, multi-step tasks. You access everything through one OpenAI-compatible API, so you don't need to manage multiple endpoints or figure out which model to call for each step. The system handles model selection and switching automatically, aiming to deliver frontier-level performance without being locked into a single vendor.

The technology behind Sakana Fugu comes from two ICLR 2026 papers: TRINITY and the Conductor. TRINITY uses a lightweight evolved coordinator to assign roles to different LLMs across multiple turns. The Conductor uses reinforcement learning to discover natural-language coordination strategies, designing communication patterns between agents. The result is a system that learns non-obvious collaboration patterns rather than relying on hand-designed workflows. Sakana Fugu comes in two tiers: Fugu (balanced performance and latency for everyday work) and Fugu Ultra (deeper agent pool for hard, high-stakes problems). You can also opt specific agents out of the pool to meet data, privacy, or compliance requirements.

Sakana Fugu is built for developers and researchers who need reliable, high-quality results on complex tasks — coding, reasoning, scientific analysis, and agentic workflows. Early users rely on it for Kaggle competitions, paper reproduction, cybersecurity analysis, and literature or patent investigations. It's not yet available in the EU/EEA while the team works toward GDPR compliance. If you're building tools that need strong reasoning and coding performance without managing multiple model APIs, Sakana Fugu offers a single endpoint that does the orchestration for you.