AI agents bring in 350,000 rubles in two months, but developer finds architecture flawed
October 23, 2026 · 2 min read
Also published in Indonesia, Deutsch, Svenska, Türkçe, Nederlands
AI agents brought in 350,000 rubles for a company in two months. The agents work on real commercial projects, handle existing repositories, and complete tasks with minimal developer involvement. The experiment proved successful from a business perspective, but the developer has significant criticisms of the system's architecture and plans to build a new one. The developer wanted to solve a specific problem: managers frequently asked developers questions about client projects, which disrupted the developers' work. The developer created Repo Chat, which allows managers to ask questions in plain language and receive answers based on the project's source code. This idea evolved into AI Workflow, which automates development tasks from assignment to merge request. The system analyzes existing code and performs small tasks independently. The developer worked with a colleague to create the system, which is now used in real projects.
The developer chose to build a custom system rather than use existing tools like Claude Code or Codex. These tools can analyze repositories and write code, but they require access to the project's context, which poses security and access control issues. AI Workflow addresses this by centrally connecting to allowed repositories and providing access through a unified interface. Managers can ask questions or assign tasks without direct access to the repository or Git credentials. The system's main advantage is its controlled use of AI development capabilities, allowing employees without direct access to source code to benefit from AI. The developer did not aim to create a better coding model than Claude or Codex but rather a managed tool for secure AI development. The system handles routine tasks that do not require complex architectural solutions but consume team time.
These tasks include fixing form behavior, changing interfaces, improving existing components, or correcting minor business logic issues. Each task requires context study, implementation, testing, and result handoff, with coding sometimes taking the least time. AI Workflow automates much of this process, leaving humans to assign tasks and accept results. On real projects, the system completed not only simple changes but also analyzed repositories, worked with multiple files, passed automatic checks, and broke down large tasks into stages. Approximately 70 percent of tasks were completed without major revisions: some were accepted immediately, while others required minor fixes through the same agent. This operational observation was convincing enough to continue using the system in commercial work.