Sakana AI's Fugu orchestrates LLM pools to survive regional shutdowns
Fugu is a coordinator model from Sakana AI that manages a pool of agents—including copies of itself—and switches between them when one becomes unavailable, outperforming single models on benchmarks.
A Tokyo-based lab founded by former Google researchers is betting that the future of AI reliability lies in orchestration, not monoliths. Sakana AI released Fugu this week, a coordinator system that routes user requests across a pool of language models and automatically fails over to alternatives when a provider goes dark.
Fugu is built on two ICLR 2026 papers from Sakana: "Trinity: An Evolved LLM Coordinator" (arXiv:2512.04695) and "Learning to Orchestrate Agents in Natural Language with the Conductor" (arXiv:2512.04388). The system runs a master model that decides which agent—commercial API, open-weight local model, or a copy of itself—handles each subtask. When one model in the pool becomes unavailable, Fugu picks the next-best option and continues without user intervention.
Sakana positions this as a hedge against the geopolitical risk that hit users of Claude Fable and Mythos, both of which were shut down abruptly in certain regions earlier this year. Benchmark results on the Fugu release page show the orchestrator outperforming any single model in its pool, including the coordinator model running standalone. The approach mirrors what translation API aggregators have done for years—routing around downtime and API degradation—but applies the same fault-tolerance logic to general-purpose LLM workflows.
Sakana suggests the pattern will become standard for local AI setups, where a lightweight coordinator calls specialized and general models on-device to match the quality of top commercial APIs without the single-point-of-failure risk.
