Untangling the code: Can an AI agent become your architecture assistant?

Refactoring a large software system can feel like trying to untangle a knot without knowing which thread to pull. Architects must identify hidden dependencies, remove unwanted coupling, extract reusable functionality and reorganize components – often without reliable design documentation. The work is valuable, but painstaking and difficult to perform manually.

This talk demonstrates how an AI agent can work directly with software architecture analysis tools to make that process faster and more systematic. Using dependency structure matrices and the Model Context Protocol, the agent can explore a codebase, detect architectural violations, identify tightly connected components and recommend practical refactoring steps. It can also evaluate architectural patterns – such as layered architecture – and suggest how existing modules could be rearranged to achieve them.

The approach is demonstrated on IsoAgLib, a 50,000-line embedded software codebase, where the agent discovered layer violations and produced concrete proposals for improvement. Early experiments achieved a 17.8 percent reduction in coupling, outperforming conventional rule-based techniques.

For software architects, this points towards a compelling future: AI that doesn’t merely generate code, but helps understand, simplify and reshape entire software systems.