Software Library Without Code Redefines Development Paradigm
- •Technologist Drew Breunig has released 'whenwords,' a software library consisting solely of specifications and tests rather than source code.
- •Developers use AI coding agents to interpret YAML conformance suites and generate functional implementations in any programming language.
- •This approach marks a shift toward specification-driven development where the primary engineering artifact is a verified test suite.
Technologist Drew Breunig has introduced "whenwords," a provocative experiment that redefines the software library by omitting source code entirely. Instead of traditional files, the repository contains detailed specifications, an instruction file for AI models, and comprehensive YAML conformance tests. This architecture shifts implementation from human developers to AI coding agents capable of interpreting complex requirements. By focusing on behavior rather than syntax, the project explores a future where software definition is decoupled from its underlying code.
The "whenwords" project utilizes an AGENTS.md file to provide clear instructions for AI models to build functional components on demand. Developers can generate implementations in any desired language by guiding coding agents through provided conformance suites. This approach leverages the capabilities of large language models to produce code that passes verified tests without manual intervention. Simon Willison, a prominent developer and technologist, noted that this experiment underscores the critical importance of robust conformance suites in an AI-driven ecosystem.
As AI coding agents evolve, the primary artifact of software engineering may transition from static codebases to well-defined test suites and specifications. This paradigm shift emphasizes specification-driven development, where the human role focuses on designing logic and verifying outcomes rather than writing syntax. The "whenwords" library serves as a proof of concept for portable, language-agnostic software that can be refactored instantly by specialized AI tools. This evolution suggests a future where software maintenance relies on refining instructions rather than debugging manual implementations.