What is alphaevolve-on-googlecloud?

AlphaEvolve on Google Cloud is a Python library and a set of runnable examples for using AlphaEvolve, an evolutionary coding agent built on Google's Gemini models. You give it a starting piece of code and a way to score how good a candidate solution is, and it repeatedly asks Gemini to propose changes to that code, tests each new version, and keeps the better ones. Over many rounds the candidates gradually improve past the original. The repository holds the client side Python package, found under src/alpha_evolve, plus several examples that show how to connect AlphaEvolve to real Google Cloud setups, ranging from a plain local Python loop with no extra infrastructure to GPU training on Google Kubernetes Engine. The examples include circle packing, which tries to fit circles into a shape while maximizing the total of their radii, tsp, which searches for shorter travel routes, signal processing, which balances several scoring goals at once, and llm fine tuning, which tunes a LoRA setup to lower its evaluation loss. The system splits into two halves. The generation half runs as a managed Google Cloud service and includes a prompt sampler, a mix of Gemini models that suggest new candidate programs, and a program database that tracks every candidate tried so far. The evaluation half is code you write yourself, since scoring depends on your particular problem. You mark the section of your seed program to be evolved with EVOLVE-BLOCK comments, write an evaluator function that returns one or more scores, and a controller loop pulls candidates from the service, runs your evaluator, and reports the results back so the next generation can improve on them. To try it, you need a Google Cloud project with AlphaEvolve already set up, meaning a Gemini Enterprise app or engine ID, along with Python 3.9 or newer, the uv package manager, and the gcloud command line tool. The quickstart example, circle packing, runs the whole loop locally without needing extra cloud infrastructure beyond calling the AlphaEvolve API itself, and can be working within about five minutes once the prerequisites are in place. The project is released under the Apache 2.0 license.