MTS Web Services has demonstrated that the era of "simple RAG" is officially over. While previous AI assistants merely mimicked interns by hunting for instructions in Confluence and Jira, the system is now entering its active agent phase. Ilya Paramoshin, Lead Engineer at MTS, describes an architecture where the neural network stops quoting manuals and starts taking action: it independently accesses logs, PostgreSQL databases, and tickets to diagnose exactly why an SMS message is stuck in transit.
In operational terms, this marks the implementation of the Model Context Protocol (MCP). MTS has transformed AI into a master key for operational data. The model now decides for itself which tool it needs at any given moment—whether that is a precise SQL query or deep log analysis. This approach allows the company to isolate resource-intensive incident investigations from trivial information retrieval. While an old-school engineer might manually piece together data from disparate systems, the agent handles this in the background via Celery workers.
Scaling beyond architectural hygiene
The scaling mechanics here are more critical than the underlying technology. Dividing the system into a RAG API and independent services for log and database management isn't just "architectural hygiene"; it is a strategy to decouple performance from data volume. Search speed no longer degrades when the model is running heavy analysis in a parallel thread. By using Open WebUI as an MCP client, engineers get a "single pane of glass" where the AI serves as a full-fledged analyst rather than a search bar.
From manual labor to quality control
The company is essentially building a pipeline where humans shift from routine fact-finding to quality control.
When the system takes over the initial ticket triage and pulls context directly from databases, support overhead stops growing linearly with the customer base. This is a rare example of the agentic approach justifying the investment, transforming technical support from a bottomless pit of payroll costs into an optimized, scalable business process.
The shift from RAG to active agents allows AI to query databases directly. Model Context Protocol (MCP) enables the AI to select its own diagnostic tools. Decoupled architecture prevents performance drops during heavy analytical tasks. Automation breaks the linear link between customer growth and support costs.