Deterministic MCP servers improve AI coding agent consistency

Alexander von Zitzewitz, founder of Hello Tomorrow and developer of SonarQube, argues that AI coding agents amplify existing software structures, leading to increased complexity if deterministic architectural constraints are absent. He cites a Stanford study observing 120,000 developers, which found an average productivity gain of only 10% from AI, with significant variance. Teams with clean code structures saw gains up to 200-300%, while others experienced slowdowns. To address this, Hello Tomorrow developed an MCP server named 'Zügel' that provides a deterministic dependency model for AI agents.

Unlike large language models that rely on grep-based search or language servers that lack full semantic dependency resolution, Zügel parses codebases in seconds to build a complete dependency graph. This allows agents to verify if new dependencies conform to a defined architecture DSL and detect cycles before code generation. In experiments with the open-source project JPOS, which has a test suite of over 4,000 tests, Claude used the MCP server to remove cyclic dependencies by introducing interfaces and inverting dependencies, maintaining test integrity. The tool supports Java, Python, and C# and enables incremental refactoring of legacy systems by identifying minimal edge removals to break cycles. Von Zitzewitz argues that architecture is the most critical asset for software longevity and that AI requires deterministic, tool-based constraints to avoid degrading code quality.