ANMA has been introduced as a tool for enforcing architectural boundaries when using AI agents for coding. The system uses YAML contracts to generate instructions (CLAUDE.md), hooks, and CI checks, allowing cheaper models to strictly follow the project structure.
What Happened
Developers have introduced ANMA, which implements external codebase control mechanisms through declarative YAML contracts. The tool automates the creation of instruction files and the integration of checks into CI pipelines. Python-based tests showed that using ANMA in combination with the Claude Haiku 4.5 model reduced module boundary violations from 13 out of 19 to 0 out of 20.
Context
When using budget LLM agents, an "architectural degradation" problem often arises, where models ignore import rules and dependencies, turning a project into spaghetti code. Traditionally, the solution was to use more powerful and expensive frontier models, but ANMA shifts control from the model's cognitive abilities to programmatic constraints (boundary contracts).
Why It Matters for the Industry
This technology changes the economics of AI agent development, allowing companies to replace expensive models with more efficient and cheaper alternatives without losing control over architectural cleanliness. This creates a foundation for the "Constrained Agent Workflows" pattern, where architectural contracts become a mandatory layer in any Software Development Life Cycle (SDLC).
Why It Matters for Users
Developers gain a way to "tame" fast and economical models, turning them into reliable assistants that do not violate project structure. This allows for the immediate implementation of budget models into workflows while maintaining strict code quality standards and reducing token costs.
What Is Not Yet Known / Limitations
There are legal risks associated with using automated agents to modify code, which require attention from intellectual property specialists.
Sources
Author
Look at AI, Editorial Team