Google's AlphaEvolve launches as a service for evolutionary code optimization

Google has opened up AlphaEvolve, a research project from DeepMind that uses artificial intelligence to automatically optimize code, as a generally available product on its Gemini Enterprise Agent Platform. This makes algorithmic optimization accessible to any Google Cloud customer who has measurable performance goals.

How it works#

AlphaEvolve functions as an evolutionary optimization agent. It takes a starting algorithm, then repeatedly generates variations using Gemini models, evaluates each candidate against a metric you define, and refines the approach until performance plateaus. The result is human-readable, optimized code rather than a black-box neural network.

The architecture preserves privacy by splitting execution: Google's AI generates candidate programs via API, but your local infrastructure—whether a laptop or supercomputer—scores them using your evaluation function and sends results back. This lets organizations with proprietary code maintain control over sensitive intellectual property.

The typical workflow follows four phases: establish a baseline algorithm and problem context, define a measurable scoring function, run the optimization loop, and deploy the winning version to production.

Early results#

Google's announcement features concrete results from customers. Klarna tested around 6,000 candidate programs over three weeks, roughly doubling ML training throughput while maintaining the exact reproducibility required in financial services. JetBrains achieved 15–20% latency improvements in IDE code completion. FM Logistic reduced warehouse picking routes by 10.4%—impressive given the baseline had already undergone production optimization. Kinaxis improved forecasting accuracy by 22% and slashed runtime by 90%. At Oak Ridge National Laboratory, the system generates optimized GPU kernels on Frontier, a world-class supercomputer.

Internally, Google has used AlphaEvolve on TPU silicon design, cutting write amplification in Google Spanner's LSM-tree compaction by 20% and storage footprint by 9%.

When it works—and when it doesn't#

A JetBrains representative captured a key insight: "Engineers still own the benchmark, review, and release decision. The search space is what gets smaller." This reflects a critical boundary: AlphaEvolve excels where the optimization target is measurable and automatable—throughput, distance, memory usage, latency, correctness. It struggles with domains requiring human judgment or where success criteria are ambiguous.