Egor Kozelsky, an engineer with 12 years of experience, has developed a framework for managing AI agents in professional software development. The system, dubbed Jarvis, addresses common issues such as loss of context across chat sessions, inconsistent project rules, and excessive token usage in large codebases.
The framework utilizes a method called progressive disclosure to manage context. Instead of loading entire project histories, agents receive a concise entry point and retrieve specific details only when required for a task. This structure is organized in a centralized directory, allowing for consistent rule application across multiple projects.
The system relies on a hierarchical directory structure where projects and groups inherit rules from parent nodes. This approach ensures that general guidelines are applied globally, while specific project constraints remain localized. The author emphasizes the use of Markdown files stored in Git to maintain version control and human readability.
To optimize agent performance, the framework implements scope routing. Agents identify the relevant project area for a given task and load only the necessary instructions. This layered instruction model prevents rule conflicts and ensures that agents follow specific architectural requirements without being overwhelmed by irrelevant data.
The author has released the base version of the tool, jarvis-installer, as an open-source project. The system is designed to support future integrations, such as RAG-based indexing or MCP connections, as project complexity grows.
