Sakana Fugu

What is Sakana Fugu?

Sakana Fugu is a multi-agent AI system that delivers frontier-level performance by dynamically orchestrating a diverse pool of the world's best models. Instead of relying on a single vendor or hand-designed workflows, Fugu learns to assemble, route, and coordinate expert agents for each task through a single API. It is grounded in two ICLR 2026 papers—TRINITY and the Conductor—which demonstrate how systems can learn to discover efficient collaboration patterns that humans might not conceive.

Who it's for

• Developers and engineering teams who need to integrate high-performance, multi-step reasoning into their applications without managing multiple model APIs.
• Organizations with strict compliance requirements that want to control which models participate in their agent pool to meet data privacy, security, or regulatory standards.
• Quality-critical workflow operators working on coding, reasoning, or complex analytical tasks that demand stronger, more reliable results than any single model can provide.

Key features

One API to access all models optimally

Fugu provides a single OpenAI-compatible API that handles model selection and switching for each task. This reduces API complexity while improving cost-performance, as the system automatically chooses the best model for every step of a multi-step workflow.

Superior performance on complex tasks

Built specifically for coding, reasoning, and other quality-critical workflows, Fugu coordinates expert agents to tackle complex tasks with stronger, more reliable results. The system dynamically assigns Thinker, Worker, or Verifier roles to adaptively delegate work across diverse domains.

Flexible agent selection for compliance

You retain control over which agents can participate in Fugu's model pool. You can opt out of specific providers or models to meet data, privacy, compliance, or organizational requirements—making it suitable for regulated environments.

What stands out

"Fugu learns to dynamically assemble agents from a pool and coordinate them through non-obvious but highly efficient collaboration patterns."

This is the core differentiator: instead of prescribing team organization or workflows based on domain knowledge, Fugu uses reinforcement learning to discover natural-language coordination strategies. The Conductor model is trained to design agent communication patterns and focused prompts that help diverse LLM pools outperform individual workers on challenging reasoning benchmarks. This learned orchestration, backed by published research, means the system can adapt and improve over time rather than relying on static, hand-crafted rules.

Worth checking out if…

You need frontier-level AI performance but want to avoid single-vendor lock-in, or if your workflows involve complex, multi-step tasks where a single model's output isn't reliable enough. Fugu is particularly valuable for teams that require flexibility in model selection for compliance reasons, or for developers who want to plug collective intelligence into their applications through a single API without managing multiple integrations. Note that Fugu is not yet available in the EU/EEA while the team works toward GDPR compliance.