Sakana Fugu Technical Report
Rohan Paul Twitter · Rohan Paul (@rohanpaul_ai) · 2026-06-28
Sakana AI's Fugu system introduces a trained orchestrator that dynamically routes tasks to specialist models and, in its Ultra variant, constructs unique multi-model critique-and-correction workflows per query rather than using fixed routing rules.
Appears in
Extraction
Topics: multi-agent-systemsmodel-routingorchestrationmixture-of-agentsai-systems
Claims
- Fugu uses a manager model trained on data—not handcrafted rules—to select the optimal specialist model for each task, including fine-grained distinctions like choosing a debugging-specialized model for code that is primarily a debugging problem.
- Fugu-Ultra can construct a distinct multi-model workflow for each individual query at inference time rather than applying a fixed pipeline.
- Most existing multi-model systems rely on simple strategies such as majority voting or static per-domain routing, which Fugu improves upon.
- The paper's central thesis is that intelligence is shifting from individual model capability to the orchestration system surrounding models.
Key quotes
The idea is that intelligence is moving from the model to the system around it.
the workflow is not fixed before the task starts, because Fugu-Ultra can design a different teamwork pattern for each question.
Fugu is different because the manager is trained from data to learn which model is actually best for each kind of situation.