Atomadic Forge: The Architecture Compiler Solving AI Code Sprawl

Why AI coding agents need an architecture compiler (and I built one)

Thomas Colvin has developed Atomadic Forge, a tool designed to enforce structural integrity in repositories managed by AI agents. The system utilizes a five-tier composition law and includes 944 tests to ensure long-term code maintainability.

Why This Matters

AI agents generate code at speeds that bypass traditional linting, leading to circular dependencies and architectural rot where utility functions call feature modules. While the resulting code executes successfully, it often lacks the structural constraints necessary for maintenance, creating technical debt that agents cannot fix without a structural feedback loop.

Key Insights

• Atomadic Forge certifies codebases with a 0-100 score and a SHA-256 receipt for structural integrity (2026).
• The system enforces a 5-tier monadic composition law where tiers only import upwards, such as pure functions in ‘a1’ only importing from ‘a0’ data.
• The tool operates as an MCP server, allowing agents in Cursor or Claude Code to use 25+ tools like certify and enforce.
• In a documented real-world test, the enforce command reduced violations from 34 to 3, achieving a final architecture score of 91.
• Atomadic Forge supports Python and TypeScript/JavaScript with a specialized ‘Deluxe’ toolset containing 35 architectural tools.

Working Examples

The 5-tier monadic composition law for file organization.

a0_qk_constants/ Zero logic. Pure data: constants, enums, TypedDicts.\na1_at_functions/ Pure stateless functions. Only imports: a0.\na2_mo_composites/ Stateful classes, clients, registries. Imports: a0+a1.\na3_og_features/ Features assembled from composites. Imports: a0-a2.\na4_sy_orchestration/ CLI, entry points, top-level wiring. Imports: a0-a3.

CLI output demonstrating the certification and enforcement process.

$ forge certify --project ./my-ai-built-app\n{\n"score": 47,\n"issues": [\n"a1 imports from a3: utils.py -> feature_pipeline.py (7 occurrences)",\n"circular dependency: auth_client <-> user_service",\n"a4 contains 340 lines of business logic (belongs in a2/a3)"\n]\n}\n$ forge enforce --apply --source ./src\n{\n"score": 91,\n"pre_violations": 34,\n"post_violations": 3,\n"auto_fixed": 31\n}

MCP configuration for integrating Forge with Claude Code or Cursor.

{\n"mcpServers": {\n"atomadic-forge": {\n"command": "forge",\n"args": ["mcp", "serve", "--project", "/your/project"],\n"env": { "FORGE_API_KEY": "your-key" },\n"type": "stdio"\n}\n}\n}

Practical Applications

• Use case: Integrating Atomadic Forge with Cursor or Claude Code via MCP to prevent utility functions from calling feature modules. Pitfall: Circular dependencies between auth_client and user_service that lead to unmaintainable code.
• Use case: Automated codebase refactoring where the enforce command fixes 91 percent of architectural violations in-place. Pitfall: Allowing business logic to leak into orchestration layers like CLI entry points.

References:

• https://dev.to/atomadic/why-ai-coding-agents-need-an-architecture-compiler-and-i-built-one-1a2f