How VibeSolve is built

VibeSolve is accessed via a simple CLI: users provide a plain-English problem description, and the system returns a fully packaged optimisation project.

VibeSolve runs the problem description through a pipeline of specialised AI agents. Each agent handles a specific part of the solver, building on the previous agent’s output. The pipeline returns a complete Java project designed to solve the optimisation problem which was originally described.

The agent pipeline

Each agent receives the accumulated ProjectManifest from all previous agents, adds its typed Delta, and passes the merged result forward.

Parser text → spec User Validator spec review Model Builder domain model Constraint Builder constraints IO Agent import / export Integrator REST + pom.xml Reviewer static code analysis Docker Validate compile + exec PASS FAIL ✓ Success project output Fixer up to max N× retry
Required agent
Optional agent
Docker validation
Agent Output Purpose
Parser Problem Specification Converts a plain-English problem description into a structured specification.
User Validator (optional) Revised Problem Specification Iteratively applies user corrections to the generated problem specification.
Model Builder Domain Classes Generates the Java domain model with Timefold annotations.
Constraint Builder Constraint Logic Implements business constraints.
IO Agent Data Layer Classes Generates JSON import/export and test data.
Integrator Main class, REST API, pom.xml, integration tests Assembles the runnable project.
Reviewer (optional) Corrected files Conducts static code analysis before Docker validation.
Fixer Corrected files Iteratively corrects compile or runtime errors reported by Docker.