Over two months, a multi-agent system generated approximately 350,000 rubles in additional revenue for the company through real commercial tasks and repository management. Against the backdrop of large-scale corporate budgets, this figure looks modest, but the metric is growing steadily. The commercial result is obvious, yet it masks the engineering dead end of the first architectural wave.
Over two months, my system of AI agents generated about 350,000 rubles in additional revenue, and that figure is only growing.
A working product and a reliable product in real-world development frequently turn out to be completely different things. The experiment proved the viability of the concept, but exposed the systemic limitations of an approach where agents write code without strict control over permissions and context. The hidden costs of supporting and scaling autonomous systems require a radical overhaul of development approaches if a business counts on a managed infrastructure rather than a one-off pilot success.
A deep analysis of inference unit economics shows that the cost of supporting autonomous agents in production can instantly consume all the margin generated by their work. When a model performs dozens of iterations to fix a single syntax error in a large codebase, the costs of input and output tokens increase exponentially. Large language model architectures have a built-in context window limitation: when attempting to process massive codebases, models inevitably lose the global architecture context, begin to "hallucinate" module dependencies, and get stuck fixing non-existent bugs. As a result, compute infrastructure expenses and manual audits of generated code outweigh the savings from routine automation, rendering the scaling of such systems financially vulnerable.
A Paid Search Engine and the Access Control Dead End
Within companies, managers approach developers five to ten times a day with routine questions about system operating logic. Developers get distracted from their tasks, and the business wastes time waiting for answers. To offload the team, Repo Chat was created, allowing employees to ask questions in plain language and receive answers based on the project's source code. The evolution of this logic led to the creation of AI Workflow, a tool that automates development from task formulation to pull request creation. A colleague of the author also participated in implementing this system.
Existing solutions like Claude Code and Codex can work with code, but they require access to the corresponding context. Granting employees direct access to commercial repositories simply to answer client questions is a critical risk to company security. Permissions must be segregated centrally rather than handed out left and right upon first request.
AI Workflow solves this problem differently, connecting to authorized repositories centrally, while employees access specific functions through a unified interface. This reduces communication overhead, but exposes the flip side: maintaining such loops consumes support resources, and the architectural limitations of current models make scaling too expensive for any pragmatic CEO.