«GigaAgent» from Sber: an autonomous AI agent that learns and rewrites its own code
Sber is bringing "GigaAgent" to market — a universal autonomous AI assistant designed for everyday and work tasks. The launch comes amid intensifying competition in the AI agent segment, where Yandex has already staked its ambitions.
The market is rapidly moving away from simple chatbots toward full-fledged digital partners capable of planning and executing multi-step processes. However, behind the loud promises of autonomy and self-improvement, there must be measurable business metrics and transparent control mechanisms — otherwise, it is just marketing.
An assistant for routine tasks
"GigaAgent" takes on calendar management, document handling, correspondence, information search, and preparation of reports and presentations. Unlike conversational chats, the new agent remembers context better and does not require oversight at every step. It accumulates feedback and adapts to an individual's personal work style.
The developers emphasize: the assistant is capable of operating autonomously in the background. It asks clarifying questions on its own when necessary and simultaneously handles multiple tasks within a single semantic framework. This is Russia's answer to foreign solutions like OpenClaw, Hermes Agent, and NemoClaw.
Three key differences between "GigaAgent" and existing analogs:
Autonomy. The agent operates around the clock in a proactive mode, honestly reports a lack of data, and can provide reasoned objections when there is an error in reasoning.
Self-improvement. It is capable of improving its own work: it reads and rewrites its own source code, recording each change as a commit in a built-in repository.
Continuity. Most AI agents operate within a single session and do not retain experience. "GigaAgent" remembers past tasks, decisions, and interaction history, resuming work after a restart while taking accumulated context into account.
Technology and pricing
At its core lies the self-improving agent project Ouroboros from researchers at the AIRI Institute. On the Terminal Bench 2.1 benchmark for Anthropic's Opus 5 model, the agent's quality reaches 86.97%.
The service can be launched through the Agents Space environment from Cloud.ru. New users receive a starting grant of 4,000 rubles for free testing. After that, usage is billed according to prices for language models and infrastructure, with the final amount depending on the volume and frequency of use.
Sber did not disclose the volume of resources allocated to the project. According to expert estimates, investments in orchestration, connectors, interface, and quality assessment systems amounted to approximately 1–3 billion rubles over a year and a half, with a team of about 100 people. For a company of this scale, that is a modest amount, especially since the base model — GigaChat 3 — has already been released under a free license. Each completed task has a cost: unlike a chatbot, the agent accesses various systems, asks follow-up questions, and makes mistakes, all of which costs money.
Earlier, Sber had already revealed plans for crypto products and AI banking for businesses, demonstrating a systematic bet on artificial intelligence.
My view: the transition from "dummy chatbots" to self-learning agents is a natural market evolution, but this is exactly where the real risks begin. The ability to rewrite one's own code requires an unprecedented level of trust and control. For now, "GigaAgent" is an ambitious technology demonstration, but investors and businesses should closely watch how Sber addresses the security and accountability issues of such systems. In the long run, the winner is not the one who loudly proclaims autonomy, but the one who offers transparent audit mechanisms and reproducible results.