Selectel and ITMO plan a platform for industrial multi-agent systems
Selectel plans to invest over RUB 1 billion in a joint venture with ITMO. We examine the proposed multi-agent platform and maturity criteria.
Selectel and ITMO University announced a joint venture for artificial-intelligence services on 6 July 2026. Selectel plans to invest more than RUB 1 billion in the project.
The announced platform
The partners describe an independent company that will develop a platform for building, deploying and operating multi-agent systems based on large language models.
The proposed product is intended to connect AI workflows to enterprise data and systems, measure quality and evolve them as business requirements change. Deployment options are expected to include Selectel infrastructure, a dedicated environment and the customer’s own site. Pricing is planned to reflect compute consumption and agent workload.
At publication time these are plans and product direction, not published evidence from a mature production deployment.
Why split work between agents
Specialised agents can separate permissions and stages such as data retrieval, policy checking, document preparation and an external-system action. Teams may use different models by cost and quality, test stages independently and replace one component without rebuilding the whole workflow.
The trade-off is coordination complexity: delegation loops, incompatible state, duplicate actions, rising cost and difficult root-cause analysis.
Production-readiness criteria
A production system needs versioned prompts, models, tools and knowledge; structured inputs and outputs; least-privilege identities; approval for irreversible operations; time, token, cost and retry limits; action logs; regression evaluation; and a stop-and-resume procedure.
Quality must be expressed as a business outcome—for example correctly completed fields, manual correction rate, cycle time and zero forbidden actions—not simply a plausible conversation.
Private deployment improves control of data boundaries but does not remove excessive permissions, hostile instructions in documents, model error or sensitive logs. Map what every agent can read, call, store and change before the pilot.
Conclusion
The Selectel–ITMO project reflects a shift from isolated chat interfaces to managed agent-automation platforms. Value will depend on reproducible quality, permission control, observability and unit economics across the complete process.
Our guide to secure AI agent architecture provides a control model. Connections to enterprise systems should follow the principles in API integration architecture.
Primary source: Selectel: AI-services joint venture with ITMO
