Sber launches GigaAgent: an agent pilot inside the Russian perimeter

On 12 August, Sber announced the launch of GigaAgent, an autonomous general-purpose AI agent. Calendars, documents, correspondence, information lookup, reports and presentations. It launches through Cloud.ru Agents Space.

What changed

The agent works without step-by-step supervision: it breaks a task into subtasks on its own, writes code when needed, and asks clarifying questions. It holds several tasks within one context and runs in the background. Underneath sits Ouroboros, an open self-improving agent project under the MIT licence, with the fork available on GitHub. The model is open as well: GigaChat 3 is published under MIT with commercial use permitted, and the repository also contains the base version straight after pretraining. New Agents Space users receive a RUB 4,000 starter grant; beyond that, billing follows consumption of models and infrastructure.

What it means

There is now an option that can be piloted without workarounds for payment and access. Companies that had planned agent automation through a foreign intermediary can postpone the decision until a measurement on a domestic platform is done. The maths should be built on cost per task, not on a subscription: the agent moves across systems, asks again and makes mistakes, and every run costs inference. The measurement is simple: take one process, run 30–50 tasks through it over two weeks, and compare against the cost of the manual process.

What is still unknown

Pricing and the billing model for corporate customers, the timeline for open access and an API, and industry restrictions. There are no public results for GigaAgent on agent benchmarks either — comparisons with foreign agents are so far made only on in-house tasks. Where the data physically sits and how the processing regime maps onto Federal Law 152-FZ has not been disclosed either, and that question is best closed before the pilot starts.