Sakana AI’s Fugu lets one API rule many AI models at once
Sakana AI has just launched Sakana Fugu, a multi-agent orchestration system that presents itself as a single model while deciding internally how to handle each request. Users send tasks to one OpenAI-compatible endpoint; Fugu then either solves the problem directly, assembles a team of expert models, or recurses with copies of itself to deliver one coherent answer.
Inside the orchestrator
Fugu is itself a language model trained to call other LLMs from a swappable pool. It manages model selection, delegation, verification, and synthesis without hard-coded workflows. When a request arrives, Fugu decides whether to act alone, farm out sub-tasks, or summon additional agents—including recursive copies of itself. From the outside, the complexity stays hidden behind a single API; inside, a coordinated system of experts does the work.
Two variants, one interface
Sakana ships two versions behind the same OpenAI-compatible API. Fugu balances strong performance with low latency for everyday coding, reviews, and chatbots; teams can opt specific agents out of its pool to meet privacy or compliance needs. Fugu Ultra targets harder, multi-step problems, coordinating a deeper pool of experts without opt-out options. The current Ultra model ID is fugu-ultra-20260615.
Built on learned orchestration
Fugu builds on two ICLR 2026 papers—Trinity and Conductor—that replace hand-designed workflows with learned coordination. Trinity assigns Thinker, Worker, or Verifier roles and adapts across multiple turns, while Conductor uses reinforcement learning to discover natural-language coordination strategies for diverse LLM pools. Together, they show systems can learn to assemble and route agents per task.
Sakana AI’s benchmarking places Fugu Ultra at or near the top on coding and reasoning tests, outperforming the individual models it orchestrates across SWE Bench Pro, TerminalBench, LiveCodeBench, Humanity’s Last Exam, and others. The orchestrator’s proprietary routing also offers a hedge against single-vendor dependency, rerouting around disruptions such as recent export controls on certain models.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.