The Map Is the App: a workflow map language for building software (github.com)

A new workflow map language (WML) has been introduced, allowing developers to declare AI software workflows as maps rather than relying on prompting agents to generate code. This innovative approach emphasizes that the map itself serves as the application, with the source code being a secondary product. By utilizing fifteen versioned primitives—such as call_llm, read_file, and loop—connected through structured contracts, WML ensures deterministic execution. This means that as long as the map is correctly parsed and the primitives are properly wired, the resulting behavior is predictable and consistent, eliminating the variability associated with model-driven code generation. The significance of WML for the AI/ML community lies in its inversion of conventional AI-coding paradigms, which often conflate model behavior with infrastructure. By treating the AI model as just one element among many primitives, WML enhances control flow, making it more robust and auditable. This approach addresses common issues like nondeterministic memory and non-restartable execution, allowing for clearer governance and enhanced reliability in AI software development. Furthermore, WML forms a critical part of the broader Symbol Grounding Framework (SGF), which aims to establish grounded machine semantics and improve the interoperability of AI systems.