Next steps

Now that you have a working Gemini Enterprise environment and have confirmed API access, you are ready to get started with the AlphaEvolve API.

Several assets and artifacts are available in the AlphaEvolve repository to help you execute your first AlphaEvolve experiment:

• AlphaEvolve Skills: Use these skills with a coding assistant to facilitate the end-to-end workflow for AlphaEvolve experiments.

• Example use case experiments (code): Explore sample code encompassing a range of optimization challenges and AlphaEvolve use cases, such as combinatorial optimization, multi-objective signal processing, and LLM fine-tuning.

• Cluster Toolkit examples: Simplify the process of deploying highperformance computing (HPC), artificial intelligence (AI), and machine learning (ML) workloads on Google Cloud. These examples serve as an accelerator for deploying custom containers for your evaluators or compiling and serving your optimized algorithms.

To begin, clone the AlphaEvolve repository or download the provided TAR file.

AlphaEvolve skills for agentic development

The AlphaEvolve skills allow you to run AlphaEvolve experiments directly from an agentic coding assistant, such as the Antigravity. Using these skills, your coding assistant guides you through the entire AlphaEvolve workflow, including configuring the experiment, running the evolutionary loop, monitoring progress, and integrating the best results back into your code.

The repository includes six specific AlphaEvolve skills, with a README.md and a SKILLS.md file provided for each:

• alpha_evolve_consultant
• alpha_evolve_experiment_design
• alpha_evolve_monitor
• alpha_evolve_orchestrator
• alpha_evolve_post_experiment
• alpha_evolve_runner

Current limitations

• Python only: The skills only support evolving Python code.

• Single code location: The system is tested to evolve one code location at a time. While context from multiple files can be included, simultaneously optimizing multiple evolve blocks is not yet validated.

Usage

1. Authenticate with the Google Cloud CLI CLI:

   gcloud auth application-default login
   

Prompt your coding assistant to optimize your code with AlphaEvolve. Example prompts include the following:

• "Use AlphaEvolve to optimize the code of this model to make it faster."

• "Set up an AlphaEvolve experiment to find a better algorithm for this scheduling function."

• "Evolve this loss function to improve convergence on this XYZ benchmark."

The assistant walks you through configuration, runs the evaluation loop, and helps you integrate the best result back into your codebase.

In addition to workflow skills, the alpha_evolve_consultant skill acts as an expert Q&A assistant for AlphaEvolve. It can help you determine whether a problem is suitable for AlphaEvolve, advise on evaluator implementation, explain runtime configuration options, and troubleshoot issues based on the official AlphaEvolve reference guide.

Note: The consultant skill is purely advisory and does not execute experiments or modify code. Invoke it by asking questions like "Is my problem a good fit for AlphaEvolve?" or "How should I design my evaluator?".

Tip: Google recommends using Google Cloud CLI 3.5 Flash as the coding assistant model. The skill works best when you provide a Markdown file that defines the problem, the objective, and relevant background details (such as constraints, evaluation criteria, or known baselines). Your code requires # EVOLVE-BLOCK-START and # EVOLVE-BLOCK-END markers around the code targeted for evolution, alongside an evaluation function. You can add these manually or ask the coding assistant for help (for example, "prepare this file for optimization with AlphaEvolve").

AlphaEvolve use cases

The AlphaEvolve repository contains the client-side Python library and example experiments in the use-cases folder. Review the README file before getting started with these curated examples:

Use Case / Level	Description	Core Learning Focus
Getting started: Combinatorial optimization (basic)	Evolve a pure-Python heuristic to pack circles into a unit square or find the shortest TSP tour using local exec() evaluation with zero cloud infrastructure.	Learn core loop mechanics (seed program → EVOLVE-BLOCK markers → evaluator → controller) and observe iterative improvement on baseline.
Multi-objective evaluation: Signal processing (Intermediate)	Evolve an adaptive time-series filter judged on over 14 competing metrics (smoothness, lag, false reversals, noise reduction, etc.) across 5 non-stationary test signals.	Learn to design weighted composite objectives, return structured insights for LLM self-correction, and protect evaluators with explicit timeouts.
Remote evaluation at scale: ML pipeline (intermediate)	Evolve a complete scikit-learn pipeline (feature engineering and model selection) scored by a containerized evaluator on Cloud Run.	Learn remote evaluation patterns using a Dockerfile, cloudbuild.yaml, and Terraform-provisioned Artifact Registry / IAM resources.
Production GPU infrastructure: LLM fine-tuning (advanced)	Evolve LoRA hyperparameters for Gemma on a function-calling dataset, trained on autoscaling L4 GPUs using a persistent RayCluster on GKE.	Learn to connect AlphaEvolve into production training infrastructure with cost controls and full observability.

Google Cloud Batch (using Cluster Toolkit)

Algorithm discovery and optimization problems vary based on the type and compute cost of their evaluators. While some evaluators run in the same Python environment as the controller, others require dedicated compilation or offloading to worker machines.

This example illustrates how to offload evaluator execution to Google Cloud Batch, using customized containers and code compilation as required. This approach lets you set up optimizations that require a complex software stack (such as running an ISV/OSS application or optimizing parts of a scientific computing software package in situ). Cloud Batch provides direct control over the required Google Cloud instance type and accelerator configurations.

The hpc-batch-ctk example provided with the AlphaEvolve assets uses:

A Pub or Sub message queue for communication between the controller and evaluators.

A Cloud Storage bucket for storing experiment results.

A JupyterLab instance to facilitate data access and visualization.

Google Cluster Toolkit is used to configure and deploy all required infrastructure components. For more information on the AlphaEvolve EAP repository for deployment, see the instructions in the hpc-batch-ctk folder located inside the README file.