Google Rolls Out AlphaEvolve to All Google Cloud Customers

Google Cloud has announced general availability of AlphaEvolve, an AI agent built on Gemini models that independently redesigns and optimizes companies' algorithms. The tool, previously tested with a closed group of customers, is now available to all users of the Gemini Enterprise Agent Platform.

AlphaEvolve is not another on-demand coding assistant. Instead of waiting for a developer's instructions, the system independently explores the space of possible solutions to a given algorithmic problem, scores them according to a scoring function defined by the client, and deploys the best ones directly to production. Google describes this as a shift from a productivity-boosting assistant to a discovery engine that finds solutions a human would not have found.

How It Works

The process consists of four steps. First, the customer defines the problem and provides a working, even if suboptimal, starting algorithm. Next, they establish an evaluation function that objectively scores the quality of candidate solutions. In the third step, AlphaEvolve's agentic mechanism generates and tests successive code variants. Finally, the highest-scoring algorithm is deployed into real infrastructure.

The technology is built by the Google DeepMind team and overseen by Pushmeet Kohli, chief scientist at Google Cloud and vice president of science at Google DeepMind. Kohli emphasizes that tools of this kind push the boundary of what is computationally achievable at all, not just speed up the existing work of engineering teams.

AI is moving beyond acting as a productivity assistant that accelerates how we work to a discovery engine that expands what we can achieve - Pushmeet Kohli, chief scientist at Google Cloud and vice president of science, Google DeepMind

Customer Results in Numbers

The list of companies that have already tested AlphaEvolve spans sectors from logistics to biopharma. FM Logistic improved warehouse routing by 10.4 percent, saving 15,000 kilometers of travel. JetBrains raised the performance of its development environment by 15 to 20 percent. PacBio, a genome sequencing company, reduced errors in genetic variant detection by 30 percent, and Schrödinger, maker of molecular simulation software, sped up the discovery of new molecules fourfold.

Coolblue, a Dutch online retailer, cut its demand forecast error by more than 5 percent over 200 optimization iterations. Advertising agency WPP saw forecast accuracy rise by 5 to 10 percent and recommendation effectiveness increase by 7 percent. Google itself uses the tool internally, among other things to design the next generation of its own TPU chips and improve data structures in its Spanner database.

What It Means for Polish Companies

For Polish technology and industrial companies, AlphaEvolve opens access to a class of tools that previously required in-house research teams focused on algorithmic optimization. The examples of BASF and Kinaxis show that the greatest gains go to companies with clearly defined, measurable operational problems, such as supply chain planning, routing, or demand forecasting.

The barrier to entry remains real, though. The tool requires the customer to define the evaluation function for solution quality themselves, meaning companies without in-house expertise in problem modeling will not benefit from it automatically. Google has not disclosed pricing for the service within Gemini Enterprise, directing interested parties to sales teams and documentation on GitHub.

AlphaEvolve does not yet support environments requiring FedRAMP certification or US Department of Defense standards, which limits its use in parts of the public and defense sectors. Google says access to such environments will be available through individual arrangements with account teams.

The launch also points to a broader trend in Google Cloud's strategy: shifting weight from generative text assistants toward specialized agents that independently run multiple cycles of trial and error within narrowly defined domains. This approach differs from direct competition with OpenAI's or Anthropic's chat models and instead targets the R&D and engineering departments of large organizations.

Sources: Google Cloud Blog (cloud.google.com), Google DeepMind Blog (deepmind.google)