Sakana Fugu makes multi-agent orchestration easier, but office AI still lives or dies on context

Sakana AI released Fugu this month. The tool wraps multi-agent orchestration inside a single API call that splits tasks, routes them across global models, and checks results.

The move lowers the engineering cost of running several models together. It also surfaces a clearer limit in office work. Orchestration speed matters less when the models lack the notes, decisions, and documents that define each project.

Fugu Ultra already matches or exceeds several closed benchmarks in engineering and reasoning. The system avoids single-vendor export controls by switching models at runtime. Those technical gains are real.

Office teams still face a different problem. They need agents that already know what was decided in last quarter's planning meeting, which clause changed in the latest contract, and why one approach was rejected. Without that record, extra agents simply generate more generic output.

Sakana Fugu office AI context therefore splits into two separate questions. One is how easily models can be coordinated. The other is whether the coordinated models have access to the actual history of the work.

Sakana positions Fugu as a way to move multi-agent systems from custom engineering projects to a standard product. Early adopters report fewer lines of glue code and faster iteration on research tasks. The same reports rarely mention production office workflows.

Orchestration alone does not create project memory

Fugu can schedule models dynamically and verify outputs. That solves the coordination layer. It does not solve the data layer that office work requires.

A typical office task references prior emails, slide decks, meeting transcripts, and shared spreadsheets. These sources sit in different tools and formats. When an agent starts without them, every new step demands the user supply the missing background.

remio stores meeting notes, documents, and browsing history in one searchable base. When a user asks for a report or a slide set, the agent already holds the context that Fugu would need supplied again each time.

The difference shows up in repeated work. A sales team that updates its quarterly forecast every month does not want to re-explain last quarter's assumptions. An agent that retains those assumptions produces consistent numbers without extra prompting.

Real projects expose the context gap

Consider a product requirements document. It draws on customer interview notes, pricing discussions from two quarters ago, and engineering constraints discussed in Slack threads. An orchestration tool can split the writing task across models. None of those models will know the constraints unless the user pastes them in.

Sakana Fugu lowers the cost of running the split. It does not lower the cost of assembling the inputs. Teams that treat context as an afterthought still spend the majority of their time gathering and explaining that context.

remio captures the inputs automatically through connectors to Notion, Linear, and browser history. The agent then works from the accumulated record rather than from a fresh prompt each session.

Benchmarks measure the wrong thing for office use

Fugu Ultra posts strong scores on engineering and science suites. Those suites test narrow problems with clean inputs. Office problems arrive with incomplete, scattered, and sometimes contradictory records.

A model that performs well on a cleaned benchmark can still produce off-target output when the underlying project history is missing. The gap is not visible in public leaderboards.

Sakana acknowledges that customers will add their own data pipelines. The company does not claim Fugu ships with institutional memory. That distinction matters for any team planning to move beyond research experiments.

What remains uncertain

Fugu is new. Adoption data is still limited to early testers. It is not yet clear how often enterprise users will connect the orchestration layer to their internal document stores or whether those connections will stay reliable at scale.

Regulators may also examine the routing logic. Dynamic model selection across borders raises questions about data residency and audit trails. Sakana has not published details on logging or access controls for those routes.

Teams evaluating the tool will need to test it against their actual project archives rather than benchmark tasks. The outcome of those tests will determine whether orchestration speed translates into faster office output.

What to watch next

Track whether Sakana or its partners release connectors that pull directly from common office tools. Adoption of any such connectors would show that users treat context as the next bottleneck.

Watch competitor responses. Other orchestration frameworks may add memory layers or tighter integrations with document platforms. Movement on this front will reveal how widely the context limitation is recognized.

Finally, monitor enterprise case studies. Published results that include cycle-time metrics on real documents or slide decks will provide clearer signals than benchmark numbers alone.

Sakana Fugu removes one engineering barrier. The remaining barrier for office AI is still the depth and accessibility of project context. Tools that address that layer will decide which orchestration improvements actually reach daily work.