Sakana AI has introduced Fugu, a platform designed to coordinate multi-agent workflows and reduce the risks tied to depending on a single AI provider in enterprise settings.
Companies encounter operational weak points when they depend entirely on one large AI API. To address these concentration risks, Japanese AI startup Sakana AI built Fugu—an orchestration-focused language model that attracts from a diverse set of models to handle complex, multi-step tasks.
Users reach this ecosystem through a single endpoint that is compatible with the OpenAI format. Fugu directs requests internally, choosing whether to answer a prompt on its own or to bring together a focused group of expert models for deeper evaluation. The system takes care of model selection, task delegation, result verification, and final synthesis behind the scenes. Engineering teams engage with what looks like a single model while a hidden layer of specialists performs the real work.
Sakana AI is aiming at the geopolitical and regulatory risks that come with procuring AI technology. Recent export restrictions targeting Anthropic models such as Fable and Mythos showed that access to particular foundational architectures can disappear overnight due to shifts in foreign policy.
Fugu acts as a safeguard against these abrupt supply chain interruptions. The platform draws on a completely interchangeable agent pool. Fugu reroutes traffic dynamically around any provider that becomes restricted or degraded, ensuring uninterrupted service. Sakana AI says this functionality delivers the resilient architecture necessary for AI sovereignty.
Fugu deployment tiers
Two tiers are offered to meet different operational latency needs.
The standard Fugu model emphasizes low latency for routine tasks and integrates into common developer tools like Codex for live coding and code review. Organizations bound by strict data governance or privacy regulations can manually exclude specific underlying models from the standard Fugu routing pool.
Fugu Ultra is built for complex, multi-step analytical challenges that call for peak accuracy. The Ultra version coordinates a broader set of expert agents for demanding work such as reproducing academic papers, conducting literature reviews, and analyzing patents.
Sakana AI reports that Fugu Ultra holds its own against top closed models like Fable 5 and Mythos Preview in scientific, engineering, and reasoning benchmarks:
The orchestration approach lets organizations tap into top-tier computing power without absorbing the vendor concentration risk or export-control vulnerabilities embedded in those closed models.
Implementation in cybersecurity
Nearly 500 early adopters put the system through its paces during an extended beta program centered on lengthy, multi-step computational workflows. With cybersecurity being a major focus for models like Claude Mythos, engineering teams put Fugu Ultra to work automating full security assessment cycles.
Human operators provided a single scoped instruction, and the orchestration engine carried out the entire reconnaissance phase on its own. The model ran cross-site scripting and SQL injection tests along with thorough authentication audits.
A cybersecurity engineer involved in the trial confirmed the model remained strictly within its operational boundaries and refrained from launching destructive actions against the target infrastructure. Fugu wrapped up the automated engagement by producing a clean vulnerability report complete with supporting evidence and precise retest instructions for human remediation teams.
The deployment showed that multi-agent routing upholds strict compliance boundaries while carrying out complex penetration testing sequences.
Software development teams also wove Fugu Ultra into their primary code review pipelines to measure defect detection rates against established monolithic tools. The orchestration engine consistently surpassed baseline models in catching logic flaws and security vulnerabilities across complex enterprise codebases.
“For code review, Fugu Ultra is significantly better than GPT-5.5. It delivers thorough answers and uncovers the bugs other tools overlook,” said a software engineer who took part in the beta rollout. “Where other tools flag around three issues, Fugu surfaced more than twenty. It has become the model I run all my reviews through.”
Automated research and persona stability
Data science teams put the system to work in a nearly fully-automated research mode. Fugu Ultra explored mathematical hypotheses, ran experimental code, interpreted failure states, and adjusted its own strategies to keep making progress over long stretches with very little human input. This capability directly tackles the operational shortcomings of single-call models that need constant human prompting to bounce back from logic errors.
Leadership at an unnamed enterprise platform company singled out long-term persona stability as a standout benefit during these extended sessions. Traditional monolithic architectures frequently suffer from context degradation and identity drift when processing long conversational histories.
“Raw output quality matches that of top frontier models, but Fugu displayed remarkably strong persona stability across long sessions, preserving its identity where other models drift,” the executive noted. “For agent-based products, that may matter more than raw benchmark scores.”
Extended benchmark validation
Sakana AI constructed the internal routing logic on the back of extensive research into learned model orchestration. The technical groundwork for the product draws from findings set out in the company’s ICLR 2026 papers, specifically the Trinity and Conductor frameworks.
These academic underpinnings enable Fugu to process requests by recognizing exactly when a task calls for delegation versus direct handling. The internal language model governs communication protocols between individual agents and structures the final synthesis of their separate computational outputs.
Validation testing against frontier AI competitors spanned complex, open-ended disciplines ranging from financial time series prediction to mechanical design. Fugu also showed strong performance in specialized physical logic tests and visual interpretation tasks, including solving the Rubik’s Cube and analyzing Japanese handwriting. The ability to excel in both quantitative financial modeling and qualitative image processing validates the effectiveness of the multi-agent orchestration strategy.
Sakana AI built the system to scale naturally as the broader AI hardware and software market evolves. Because the product depends entirely on learned orchestration logic rather than rigid operational rulesets, it automatically gains from third-party innovations. Sakana AI intends to keep expanding the available pool of expert agents.
The engineering team will fold newly released open-source tools and proprietary Sakana AI models into the routing pool as they become available. Both the standard Fugu and Fugu Ultra models are available to enterprise clients today.
See also: SAP and Google Cloud deploy agentic commerce architecture
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
