Sakana AI released Sakana Fugu, a multi-agent system that packages several large language models into one deployable unit. Unlike traditional approaches that require separate model instances, Fugu orchestrates multiple frontier models dynamically to handle complex, multi-step tasks through standard API endpoints.

The system uses a compact coordinator model optimized with evolutionary strategies, delegating roles to a pool of LLMs without weight merging or shared architectures. A Conductor model, trained via reinforcement learning, designs agent communication topologies and writes targeted instructions for each worker LLM. This approach lets different models collaborate turn by turn, discovering coordination strategies that outperform individual models.

Fugu demonstrates superior performance compared to anonymized baselines equivalent to Gemini 3.1 Pro, Opus 4.8, and GPT 5.5. The system continuously updates its model pool and retrains coordinators as new frontier models emerge, maintaining competitive advantage through ongoing optimization rather than static architectures.

Developers can install Fugu with a single command on Ubuntu and macOS, integrating it directly into Codex environments. The one-model delivery simplifies deployment while preserving the benefits of multi-agent collaboration for tasks requiring diverse reasoning capabilities.