Use case implementation guidelines

Once you have decided that an optimization use case is a good fit for AlphaEvolve, use the following steps to implement the use case and run the required experiments:

  1. Environment setup and compliance:

    1. Configure environments: Set up the Gemini enterprise or discovery engine environments and AlphaEvolve service accounts.

    2. Compliance check: Validate any infosec, risk, and compliance constraints applicable to the environments that are relevant to your use case and industry.

  2. Initial onboarding: Familiarize yourself with how AlphaEvolve works by running the tutorial colabs.

  3. Define experiment parameters: Define your experiment parameters and optimization objectives based on your use case requirements:

    1. What is your optimization objective?
    2. What programming language should be used?
    3. What is your design space and what are your constraints?

  4. Experiment design and configuration:

    You can approach this stage through either manual configuration or an automated agentic path:

    Option A: Manual experiment design and heuristic configuration:

    1. Formulate the initial seed program, add context, and tag with appropriate #EVOLVE-BLOCKS.

    2. Design evaluator metrics and implement the evaluation harness accordingly.

    3. Configure heuristic hyperparameters and start the experiment.

    Option B: Alternative agentic path: Use your agentic coding framework of choice (Gemini CLI, Antigravity, and so on) along with the AlphaEvolve skills files to design your experiment,build the evaluation harness, and configure the heuristic parameters.

  5. Execution and analysis:

    1. Monitor: Observe your experiment closely and check for any convergence issues that might arise.

    2. Analyze: Review the results thoroughly.

  6. Iterate:

    1. Iterate on the initial experiment as needed by tweaking the context, evolve blocks, and evaluation metrics based on the results from the initial run.

    2. Most interesting AlphaEvolve use cases will require more than one heuristic run to get optimal results.