September 23, 2026
Terminology checks in the IDE: SberTech agent works with 750 rules
On September 23, SberTech unveiled an agent for technical writers: the Python logic took a day to build and another week to debug.
The team first connected Continue in VS Code to its internal Qwen3-32b and loaded a glossary with prompts into Langflow. It then converted the glossary from Markdown to YAML and moved the logic to Python with FastMCP and openai.
The threshold determines quality. The glossary contains more than 750 terms, so the team selected a match score of 0,85–0,90. At 0,99, the agent misses less obvious terms, while a lower threshold brings in extra ones. In the IDE, users can check a single file or a documentation set, but formatting sometimes shifts after the model makes edits.
The team is developing an MCP server with an API to run this logic as a script, a local MCP server, or through an API.
Source
