The Intellect Foundation held the first scientific and practical seminar on AI agents for business and industry
On July 2, the Faculty of Artificial Intelligence of Moscow State University hosted the scientific and practical seminar of the Intellect Foundation "AI agents: types, trends, scientific outlook and applied application". The seminar was attended by both AI researchers and representatives of major Russian companies, including technical and IT directors, digital transformation managers, and experts in agent systems.
presentation of the report The seminar was opened by Anton Konushin, Program Director of the Intellect Foundation, Candidate of Physico-Mathematical Sciences, Deputy Dean of the Faculty of Artificial Intelligence for Scientific Work. In his introductory report, he presented the general concept of a regular AI trends workshop.
"AI technologies are developing exponentially, and many companies around the world are piloting implementations - but then they discover unexpected effects of such implementations. There is a noticeable gap between the potential of AI methods and the practice of their application in business. In this regard, one of the goals of our seminar is to discuss the use of AI and find new ways to implement it as effectively as possible.", - noted Anton Konushin. He outlined the general context of AI development in the world and in Russia, where access to advanced models is currently limited, and therefore many solutions are based on further training of open-source models.
With the report "How industrial engineers became vibe coders. A brief history of AI trends" was delivered by Mikhail Sukhorukov, a research programmer at ANO "Neuro-Knowledge". He talked about how, since the release of ChatGPT and the spread of the Transformer architecture, dialog models and prompt engineering have changed the approach to working with AI, making the quality of the response directly dependent on the wording of the request, context and role. Mikhail showed the evolution from simple chatbots and classical machine learning to bundles of RAG, LLM workflows and agent circuits, where language models are embedded in classical code and control data processing routes. Special attention in the report was paid to the transition from "manual" manufacturing to the design of the AI environment and context: context engineering, the design of skills and configuration files for agents (Agent Skills), as well as the practices of vibe coding and agentic engineering. Mikhail showed how modern AI agents go beyond the development environment and become an integral part of the user's "digital life", integrating with mail, browser and files, and how this shifts the role of a person from writing code directly to setting tasks, monitoring and designing a secure environment for change.
Vera Krasnobaeva, programmer-researcher at the NGO "Neuro-Knowledge", in the report "Agent-based solutions: what they are" spoke about approaches to building multi-agent systems. She started with the basic definition of an agent as a performer acting on behalf of a principal, and showed that an AI agent is a bundle of LLM, autonomy, memory, tools, and a perception-thinking-action chain, unlike a chatbot that only works in dialogue with a human. Vera further presented six architectures of agent-based solutions: workflow/graph for sequential logic, orchestration with a central coordinating agent and subagents, handoff scenarios with the transfer of dialogue or tasks to "narrow specialists", competitive "tournaments of ideas", contours of reflection and self-criticism, as well as a blackboard architecture with a common "board" of events. Using examples from the practice of international companies, she showed when it makes sense to choose one approach or another, what risks different architectures carry, and how architecture affects not only the use of AI, but also the distribution of responsibility, context, and control between parts of the system.
In the second block of the seminar, Anton Konushin made a presentation on "Evolutionary methods for the development of programs, methods and agents." He talked about the idea of AlphaEvolve, an agent-based system that uses LLM to find and evaluate code modifications that allow for continuous acceleration of model learning. Using the example of AlphaEvolve, it was shown how the evolutionary approach helps to improve system performance step by step, up to freeing up significant computing resources and reducing training time without radically redesigning the architecture. Anton Konushin spoke about GigaEvo, an open implementation of the AlphaEvolve system, as well as its related approach to creating agent chains through the Collaborative Agent Reasoning Library (CARL), where each step is set in a structured manner, which allows LLM to better understand tasks and offer new-level solutions. He gave examples of evolving agents capable of rewriting their own components, and spoke about the Maestro system, a platform for generating and evolving agents and their configurations. In conclusion, Anton noted that in Russia, the AI Institute is forming and actively developing an open ecosystem of evolutionary methods for building and developing agents.
Ivan Matveev, Head of Digital transformation at Ingosstrakh Joint Stock Company, in his report "When ERP and RPA meet Agentic OS" showed that the main challenge today is not creating one agent, but scaling dozens and hundreds of agents without loss of manageability. If each agent is developed in isolation, with its own security and without a general audit, the company gets not order, but "new chaos" - all the more dangerous because agents make decisions autonomously. Curbing Chaos is an iMan platform that implements the concept of Agentic OS ("operating system for agents"). This is a single environment where agents go through a full lifecycle: from creation by a business user through no-code or by a developer through low‑code to testing in an isolated evaluation loop, then industrial operation with centralized rights, audit and integration through standard protocols (MCP, A2A). iMan does not replace existing ERP, RPA and CRM, and allows agents to conduct them - to call them as tools within the framework of approved rules.
Using the example of a case on finding and attracting clients, Ivan Matveev demonstrated how an agent, having received a goal ("to select news, enrich data and decide whether to start a lead"), independently performs all the steps: analyzes the news stream, accesses an external service (Contour.Focus) for data, checks the history in CRM and decides whether to create a lead or create an interest. In the pilot mode, the selection accuracy reached 96%, and employees switched from manual processing to criteria management and dispute control.
Today, agents are already working on the platform in several departments: the legal department (speeding up the formation of excerpts from court documents by 50%), the Data Science center (reducing the time to retrain ML models from 5-11 days to 1 day) and others. At the same time, all the logic of agents, prompta and skills are stored in an open format, the data does not leave the contour of the company, and the business model remains an asset of the customer - the platform does not "sew up" ready-made scenarios, but provides the foundation for self-configuration for any processes.
"The winners are not those who launch one agent first, - Ivan Matveev summed up.
The winners are those who are the first to learn how to connect agents to their systems in a large and manageable way. iMan is a platform that provides this opportunity without requiring to rewrite the existing infrastructure."
