AlphaEvolve

// Overview

AlphaEvolve is an evolutionary coding agent for general-purpose algorithm discovery and optimization. You provide a seed program that already works and a scoring function that says what "better" means; AlphaEvolve uses Gemini to propose code changes, evaluates each candidate against your metric, and evolves the population over many generations toward higher-scoring programs. Google's documentation frames it as a fit for algorithmic discovery, mathematical search and combinatorial optimization — including NP-complete and NP-hard problems where the win comes from redesigning the algorithm, not tidying the code.

It is deliberately not a general code assistant. The documentation is explicit that AlphaEvolve is not for basic code generation, linting or cleanup: it starts from functionally correct code that needs a better score, and it needs a large space of functionally correct alternatives to search through. The trade-off is compute — a run is a long loop of generate, evaluate, select, so cost scales with how many candidates you are willing to evaluate.

The service runs inside your own Google Cloud project through the Gemini Enterprise Agent Platform. The public repository is the client side: a Python library (`alpha_evolve`) plus runnable examples that wire AlphaEvolve into real Google Cloud environments, from a pure-Python loop with no infrastructure at all up to GPU training on GKE. The library is Apache-2.0, but running it requires a Google Cloud project with AlphaEvolve provisioned, which gives you the Gemini Enterprise App/Engine ID the examples expect.

// What it does

• Evolutionary search over programs: Gemini proposes code changes, each candidate is evaluated against your scoring function, and the population evolves across generations
• You define the objective — the agent optimizes whatever metric your evaluator returns, rather than a fixed notion of code quality
• An open Python client library (`alpha_evolve`) with runnable examples spanning local evaluation, Cloud Run and GPU training on GKE
• Parallel evaluation via PARALLEL_EVALUATION and WORKER_CONCURRENCY settings, so candidates are scored concurrently
• Runs inside your own Google Cloud project through the Gemini Enterprise Agent Platform
• AlphaEvolve Skills for driving runs from an agent environment, documented alongside the library

// Getting started

The Circle Packing example runs the full evolution loop with local Python evaluation and no cloud infrastructure beyond the AlphaEvolve API — the fastest way to watch the loop work. You need Python 3.9+, uv, the gcloud CLI, and a Google Cloud project with AlphaEvolve provisioned.

git clone https://github.com/Google-Cloud-AI/alphaevolve-on-googlecloud.git
cd alphaevolve-on-googlecloud
uv pip install -e ".[examples]"

cd examples/circle_packing
make setup

make auth

make run

PARALLEL_EVALUATION=True
WORKER_CONCURRENCY=8

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Search for a faster implementation of a hot algorithm when you can write an automated benchmark that scores each candidate
• Attack combinatorial optimization problems — packing, scheduling, routing — where many functionally correct programs exist and only the score separates them
• Explore mathematical constructions by encoding the construction as a program and the objective as a metric
• Tune a kernel or training configuration on GKE by letting the loop propose and measure variants instead of sweeping by hand

// How AlphaEvolve compares

AlphaEvolve alongside other open-source autonomous coding agents tools AI/TLDR tracks, ranked by GitHub stars.