A Dynamic Multi-Agent Architecture: Preset-Free and Self-Organizing
Japanese AI research company Sakana AI has officially launched its flagship commercial product, the Sakana Fugu multi-agent orchestration system, and opened applications for early beta testing. This system is not a single model but an intelligent framework designed to dynamically coordinate multiple AI "Workers."
Product Line Positioning
The Fugu series offers two configurations: the Sakana Fugu Mini, optimized for low latency, and the Sakana Fugu Ultra, designed for demanding and complex tasks. Both are delivered via a single-model API compatible with the OpenAI format, greatly simplifying integration for developers.
Core Technology: Autonomous Learning and Dynamic Orchestration
The system's design philosophy stems from the team's Trinity and Conductor research papers presented at ICLR 2026. At its core is a lightweight language model, but it acts as a "conductor," not a "soloist."
- Dynamic Task Allocation: The system discards the traditional multi-agent approach requiring manually pre-defined team roles or fixed pipelines. It automatically calls suitable models from a Worker pool and dynamically allocates tasks based on the perceived difficulty of the received job.
- Adaptive Recursive Calling: Fugu possesses unique "test-time scaling" capability. The model can read its own previous outputs as context, autonomously identify flaws or deficiencies during runtime, and automatically initiate correction workflows.
- Controllable Compute Axis: Users can treat "recursion depth" as a tunable parameter during inference. This allows developers to flexibly control the system's "thinking" intensity based on the need for result accuracy versus response speed.
Performance Benchmarks: Surpassing Top-Tier Single Models
According to official evaluation data released by Sakana AI, Fugu Ultra has achieved breakthrough scores on several challenging benchmarks, outperforming current leading flagship single models.
Benchmark Results
- GPQAD (General Purpose Question Answering & Reasoning): Score 95.1
- LCBv6 (Logic & Commonsense Benchmark v6): Score 93.2
- SWEPro (Software Engineering Professional Test): Score 54.2
In the above tests, Sakana Fugu Ultra's overall performance surpassed that of models like GPT 5.4, Gemini 3.1, and Claude Opus 4.6. Its architectural advantages of multi-agent collaboration and dynamic error correction are particularly evident in hardcore reasoning and coding tasks.
The launch of the Fugu system marks a critical step in moving multi-agent collaboration from theoretical research to large-scale commercial application. Its paradigm of "dynamic orchestration and autonomous optimization" may bring new perspectives to developing complex AI applications.