The authors introduce Sakana Fugu, a family of orchestrator models that are themselves language models trained to understand user queries and dynamically build agentic scaffolds—teams of LLM agents—to solve tasks. These scaffolds allow Fugu to access capabilities beyond any individual LLM, achieving state-of-the-art performance on challenging benchmarks including SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. Two models are released: Fugu (balancing performance and latency) and Fugu-Ultra (prioritizing quality on hard problems). The training paradigm includes large-scale fine-tuning, evolutionary algorithms, and , with infrastructure details provided. This work demonstrates that collective intelligence from multiple specialized LLMs can reach new frontiers of AI capability.
Orchestrator models dynamically combine specialized LLMs into collective intelligence
The authors train a language model to understand user queries and generate adaptive agentic scaffolds, outperforming individual models on coding, reasoning, and terminal tasks.
AI Startup
Yujin Tang · Edoardo Cetin · Jinglue Xu · Qi Sun · Stefan Nielsen · Vincent Richard · +8 more
Sakana AI
Research Digest··2 min read
Thread:Agent Harness Optimization
In one line
Fugu orchestrator models combine multiple LLMs into adaptive agent teams to achieve state-of-the-art results.
Before you read this
Three ideas this paper builds onmulti-agent architectureA system where multiple AI agents collaborate, each handling specialized roles, communicating and coordinating to complete complex tasks.99 papers →agentic inference-time scalingSpending more computation during inference, often by running multiple agent iterations or expanding context, to improve model output quality.18 papers →model routingDirecting each query or subtask to the most suitable model to optimize cost, accuracy, or quality.75 papers →
Read it as
The authors present Sakana Fugu, a family of orchestrator language models that generate dynamic, query-adaptive agentic scaffolds to harness the specialized capabilities of multiple LLMs.
The authors trained two orchestrator models (Fugu and Fugu-Ultra) using a combination of large-scale fine-tuning, evolutionary algorithms, and .
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.