Software engineering is entering a phase transition. The traditional loop—human reads spec, human writes code, human runs tests, human fixes bugs—is being replaced by something structurally different: an autonomous agent loop where the human authors intent and reviews outcomes, while an orchestrated system of AI agents handles decomposition, implementation, verification, and retry. eval-loop is a working implementation of this paradigm.
This is not a copilot. It is not autocomplete. It is a Spec-Driven Development (SDD) pipeline engine that takes a product proposal—outcomes and success criteria written in product language—and compiles it into dependency-ordered engineering changes, implements each one through fresh Claude Agent SDK sessions, and runs a closed assess-fix-verify loop until every change passes both deterministic tests and LLM-based evalua