Agents Should Not Be Written by Developers. Here's Why
Cloud.ru
The AI automation team at Cloud.ru found that classic AI automation stopped scaling because the bottleneck was not the LLM but the engineers' domain knowledge, which is hard to transfer. They shifted from writing agents themselves to removing barriers so engineers could create their own agents, resulting in unexpected use cases and a community called 'AI Vanguard'.
At Cloud.ru, the AI automation team built an internal platform for customer support with AI models, automatically classifying requests, summarizing communications, and routing tickets. However, progress stalled because each new scenario required weeks of immersion into specific services, and there are over 200 services. The team realized the bottleneck was not the LLM but the engineers' domain knowledge, which is hard to transfer. Citing Microsoft's Work Trend Index 2026 and BCG's AI at Work 2026, they concluded that the goal is to enable engineers to build their own agents using their expertise. They started by evangelizing, removing technical barriers by providing local models, guardrails, a safe environment, and MCP servers. Unexpectedly, engineers created agents for diagnostics, internal APIs, log analysis, and Kubernetes checks. To sustain momentum, they built a community called 'AI Vanguard' with leaderboards, internal currency, and achievements, and are now focusing on scaling this practice to 100 people, with details to follow in a next article.
- Abbreviations
- LLM = Large Language Model — Большая языковая модель
- MCP = Model Context Protocol — Протокол контекста модели
- API = Application Programming Interface — Интерфейс программирования приложений
- KPI = Key Performance Indicator — Ключевой показатель эффективности
Source: Habr — хаб ИИ —
original