VkusVill launched AI agents as digital employees on the OpenClaw platform

VkusVill shared how it builds a team of AI agents following the principle of hiring employees: each agent receives a role, undergoes training, and is retained in the system — like a regular specialist, only digital.

Agent as a Work Unit

VkusVill's AI Expertise Center has deployed AI agents on the OpenClaw platform. The approach differs fundamentally from point-wise model application: instead of one-off calls to an LLM, the team builds a full-fledged "onboarding" — each agent receives a clear role, scope of responsibility, and structured context about the company, its products, and processes.

OpenClaw allows assembling agents from tools, memory layers, and system prompts, then linking them into working chains. For a retailer with thousands of SKUs this is critical: tasks vary from processing customer requests to monitoring assortment and supply chain, and each requires a specialized agent with its own expertise.

The architectural principle the team considers key: an agent does not work in isolation. It must be able to delegate tasks to other agents when it exceeds its competence, and correctly receive results from neighboring "colleagues" — otherwise the entire chain breaks down.

Three Main Challenges

According to Sabina, product owner of VkusVill's AI Expertise Center, launching digital employees comes down to three tasks:

• Hiring — select the right agent architecture for a specific function. A mistake here leads to excessive complexity and unstable behavior already in production

• Training — provide the agent with live context about the company: not just a set of documents, but structured and regularly updated memory about products, processes, exceptions

• Retention — preserve accumulated behavior and "experience" during base model updates or environment configuration changes

The last point turned out to be the most non-obvious. When a provider releases a new model version, agent behavior changes — sometimes imperceptibly to developers, but significantly for real business processes. VkusVill's solution: versioning of agent configurations and a set of regression tests that automatically run with every dependency update.

Next Step — Customer Agents

While agents work within the company as colleagues, VkusVill is already designing an external scenario: agent as a customer. At the end of 2025, the team published material about an experimental MCP server — an API layer through which an external agent could independently browse the catalog, compare items, and form an order.

"While we prepare for customer agents, we will tell you how agents become our colleagues," —

Sabina, product owner of VkusVill's AI Expertise Center.

An updated version of the MCP server is expected in spring 2026. This is part of a broader trend: retail is beginning to design interfaces not only for humans, but also for autonomous agents. For business this means the need to prepare machine-to-machine APIs in advance: catalog structure, product data formats, authorization protocols for customer-agents.

What This Means

VkusVill demonstrates a mature view of industrial AI implementation: an agent is not a script or chatbot, but a work unit with its own life cycle. HR terminology — hiring, training, retention — now applies to software systems. Companies that build these processes now will occupy a strong position when agent architectures become an industry standard.