Sakana Fugu is a multi-agent orchestration system developed by Sakana AI that dynamically coordinates pools of frontier large language models (LLMs) to achieve state-of-the-art performance on specialized benchmarks. The system represents an approach to leveraging multiple LLM instances in coordinated workflows, with particular emphasis on complex reasoning and domain-specific problem-solving tasks.

Sakana Fugu operates as a multi-agent orchestration platform that coordinates multiple frontier-class LLMs rather than relying on a single unified model. The system dynamically routes tasks and coordinates responses across a pool of LLMs, allowing specialized models or configurations to focus on particular problem domains. This orchestration approach enables the system to distribute computational load and leverage the distinct strengths of different LLM variants or sizes ¹⁾

The dynamic coordination mechanism allows Sakana Fugu to allocate resources adaptively based on task complexity and requirements. Rather than applying a single inference path, the system can route different components of problems to specialized agents, combine outputs intelligently, and refine results iteratively. This architecture contrasts with single-model approaches by distributing the inference process across multiple capable systems.

Sakana Fugu achieved state-of-the-art (SOTA) results on two significant evaluation benchmarks: SWE-Pro and GPQA-Diamond. The SWE-Pro benchmark evaluates software engineering capabilities, including code generation, debugging, and complex technical problem-solving tasks. GPQA-Diamond represents a challenging set of graduate-level questions spanning multiple scientific domains, requiring deep domain knowledge and sophisticated reasoning ²⁾

These benchmark achievements suggest the orchestration approach effectively improves performance on specialized technical reasoning tasks. The multi-agent design appears particularly suited to domains where decomposition, specialized expertise, and iterative refinement provide advantages over single-model inference.

As of May 2026, Sakana Fugu entered public beta with tester applications available to evaluate the system in real-world deployment scenarios ³⁾. The public beta phase allows external developers and organizations to access the orchestration system and provide feedback on performance, usability, and integration capabilities.

The beta testing window represents an opportunity for developers building applications requiring high-performance reasoning on specialized tasks. Organizations working with software engineering automation or scientific question-answering could benefit from early access and evaluation of the system's capabilities.

Sakana Fugu exemplifies a broader trend in LLM system design toward multi-agent architectures. Rather than scaling individual model size indefinitely, orchestration approaches distribute specialized capabilities across multiple agents that coordinate toward shared objectives. This design pattern reflects insights from distributed systems architecture and agent-based AI, where coordination mechanisms can achieve emergent capabilities exceeding individual components ⁴⁾

Multi-agent systems enable specialization, where different agents develop distinct competencies and knowledge distributions. The coordination layer manages task decomposition, agent selection, output combination, and result validation. This approach has shown promise for reasoning-intensive applications where breaking complex problems into manageable subtasks improves solution quality.

The demonstrated performance on SWE-Pro suggests Sakana Fugu is particularly effective for software engineering applications including code generation, bug identification, and technical documentation. Organizations developing AI-assisted development tools could leverage the system for tasks requiring both broad programming knowledge and deep reasoning about complex technical problems.

GPQA-Diamond performance indicates capabilities for scientific reasoning and knowledge synthesis across multiple domains. Academic institutions, research organizations, and knowledge-intensive applications may find value in deploying multi-agent systems for literature synthesis, hypothesis evaluation, and cross-disciplinary problem-solving.

• Sakana Conductor
• Sakana AI
• Sakana KAME
• Frontier Models Ensemble
• Small Language Model Agents