ANMA, a new open-source tool, uses module contracts to keep AI coding agents (like Claude) inside architectural boundaries, cutting boundary violations from 68% to 0% in benchmarks for cheaper models.

What happened: ANMA Labs released ANMA, a tool that converts plain YAML module contracts into generated documentation and enforcement hooks. In a Python benchmark, Claude Haiku 4.5 violated declared module boundaries in 13 of 19 runs without ANMA; with ANMA, it violated boundaries 0 times across 20 runs (Fisher's exact p < 0.0001). TypeScript showed similar results in a follow-up study (control 18/20 violations vs. ANMA 0/20, p < 0.00001); Go results were directional and significant (10/30 → 0/30, p = 0.0004).

Why it matters: Teams running cheaper or faster AI coding agents—where cost, speed, and reliability matter more than frontier-model performance—currently cannot guarantee their code stays within declared architectural boundaries without manual oversight. ANMA provides both guidance (via generated architecture documentation the agent reads) and enforcement (via hooks that block disallowed edits before they land in CI or pre-commit), offering a measurable safety layer for budget-conscious automation.

What to watch: ANMA is lightweight (~800 lines, Apache-2.0, one small dependency), supports Python, Go, and TypeScript, and includes enterprise features like drift detection and incremental adoption modes. The tool is available now via pip install anma[tach]; the full benchmark methodology and honest limits are documented in docs/BENCHMARKS.md.