Sakana Fugu: a multi-agent system delivered as one model

Sakana AI launched Fugu, a multi-agent system delivered as a single model that dynamically orchestrates frontier models to tackle complex multi-step tasks. The system is accessible via the Sakana API and can be installed into Codex with a single command, achieving superior performance through dynamic coordination of diverse models.

Sakana Fugu https://sakana.ai/fugu/ is a multi-agent system delivered as one model. Fugu dynamically orchestrates frontier models to tackle complex, multi-step tasks. You can access the multi-agent system as a single LLM through the Sakana API https://console.sakana.ai/get-started , which supports both Chat Completions and Responses endpoints. To quickly get started, you can install Fugu into Codex with a single command: curl -fsSL https://sakana.ai/fugu/install | bash Then launch it with: codex-fugu See the command reference /SakanaAI/fugu/blob/main/docs/commands details.md for additional flags and options. The one-line install supports Ubuntu and macOS. On Windows, or if the install does not complete, follow the guide here https://console.sakana.ai/get-started manually-setting-up-codex . Sakana Fugu achieves superior performance by dynamically coordinating and orchestrating a diverse pool of powerful models. For evaluation details, check our technical report /SakanaAI/fugu/blob/main/Fugu technical report.pdf . These results reflect our June 2026 evaluation. As new frontier models are released, we continuously update our model pool and retrain our coordinators to maintain Fugu's performance advantage. These examples compare Sakana Fugu models with three frontier baselines: Gemini 3.1 Pro high , Opus 4.8 max , and GPT 5.5 xhigh . To keep the focus on behavior rather than brand-by-brand attribution, the baselines are anonymized as Model A, Model B, and Model C in each description. The mapping is intentionally not fixed across examples. Sakana Fugu is based on two papers published in ICLR 2026. A compact coordinator model, optimized with an evolutionary strategy, delegates three roles to a pool of LLMs turn by turn, letting them collaborate without weight merging or shared architectures. | A Conductor model, trained with reinforcement learning, designs agent-to-agent communication topologies and writes targeted instructions for each worker LLM, discovering coordination strategies that outperform any individual model. | Since publication, we have made several enhancements. The full technical report is available here /SakanaAI/fugu/blob/main/Fugu technical report.pdf . Please contact us at https://fugu.sakana.ai https://fugu.sakana.ai for issues or bugs while using Sakana Fugu. If you use Sakana Fugu in your research, please cite our technical report: @misc{fugu2026sakana, title={Sakana Fugu Technical Report}, author={{Fugu Team, Sakana AI}}, year={2026}, eprint={2606.21228}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2606.21228}, }