Sakana AI’s Fugu multi-agent API aims to limit vendor lock-in

Sakana AI launched Fugu, an OpenAI-compatible orchestration system that exposes a single API endpoint and routes requests across a swappable pool of specialist models to limit vendor lock-in and work around export controls.

An internal orchestration language model decides whether to answer a prompt directly or to assemble a team of expert agents. The system handles model selection, task delegation, result verification and synthesis so engineering teams interact with what appears to be one model while multiple specialist models perform the underlying computation.

Sakana developed the product to address geopolitical and regulatory sourcing risks. The company cited recent export controls that affected Anthropic models as an example of how access to specific architectures can be cut off. Fugu’s swappable agent pool and dynamic routing are designed to maintain service continuity by shifting traffic away from restricted or degraded providers.

Fugu is offered in two tiers. The standard tier is tuned for low latency and integrates with developer tools for live coding and code review; organisations can opt specific underlying models out of the routing pool for data-governance reasons. Fugu Ultra coordinates a larger set of expert agents for multi-step analytical tasks such as reproducing academic papers, conducting literature reviews and analysing patents.

Sakana reported that Ultra performs competitively with closed models including Fable 5 and Mythos Preview on scientific, engineering and reasoning benchmarks.

Nearly 500 early users tested Fugu during an extended beta focused on long, multi-step workflows. In cybersecurity use cases, teams ran automated assessment cycles where a single instruction triggered reconnaissance, cross-site scripting and SQL injection checks and authentication reviews. The orchestration engine produced vulnerability reports with verifying evidence and exact retest steps for remediation.

A participating cybersecurity engineer confirmed the model stayed within operational parameters and did not initiate destructive actions against target infrastructure.

Software development teams integrated Fugu Ultra into code review pipelines and reported higher defect detection rates. A software engineer involved in the beta reported that Fugu Ultra surfaced more than twenty issues on some reviews compared with about three flagged by other tools.

Data science teams used Ultra in near-automated research modes. Fugu Ultra explored mathematical hypotheses, executed experimental code runs, interpreted failure states and revised strategies over extended sessions with limited human input. An executive at an unnamed platform company noted that output quality matched top frontier models and that the system maintained a consistent persona over long interactions.

Sakana built the routing logic on research published at ICLR 2026, citing the Trinity and Conductor frameworks as technical foundations. The internal language model evaluates when delegation is necessary and designs the communication protocols between agents before synthesising outputs.

Validation testing covered tasks from financial time-series prediction to mechanical design and visual interpretation, including solving a Rubik’s Cube and Japanese handwriting analysis. Sakana stated Fugu’s learned orchestration can incorporate new third-party models automatically and that the firm plans to add more open-source and proprietary agents over time. Both Fugu tiers are available to enterprise clients today.