Google Created AI Agent AlphaEvolve to Automate Algorithm Design

DeepMind introduced AlphaEvolve — an AI agent that automatically designs and optimizes complex algorithms. The system combines the creative capabilities of Gemini large language models with automatic evaluation functions and an evolutionary algorithm for successive improvement.

How Evolutionary Search Works

AlphaEvolve uses an ensemble of Gemini models to divide tasks. Gemini Flash — a fast and efficient model — is responsible for generating the maximum number of ideas, exploring a wide space of possible solutions. Gemini Pro, Google's most powerful model, adds critical depth, offering thoughtful algorithmic solutions that take into account complex mathematical principles.

Each candidate is source code that implements the proposed solution. The system runs this code and verifies the results using automatic evaluation metrics. This provides an objective, quantitatively measurable assessment of the accuracy and quality of each algorithm.

This approach is particularly powerful in domains where progress can be clearly and systematically measured — in mathematics and computer science. Based on these evaluations, the evolutionary algorithm selects the best solutions, which become the foundation for the next generation of candidates.

The system accumulates all algorithms ever discovered in a database and uses this database to generate new proposals. The process iterates, ideas improve, and results gradually get better.

Practical Results at Google

AlphaEvolve has already been deployed to production at Google and demonstrates impressive results:

• Data Center Optimization — the system discovered a simple yet effective algorithm for scheduling computational tasks in Borg, Google's massive data center orchestration system. This algorithm has been running in production for a year and saves an average of 0.7% of Google's computing resources worldwide

• Chip Design — AlphaEvolve proposed rewriting a critical matrix multiplication chain, removing redundant bits in key operations

• Accelerated Model Training — algorithms discovered by the system have been incorporated into the training process of modern LLMs, including Gemini and Gemini Flash themselves

• Mathematical Discoveries — the system found new solutions to several open mathematical problems

Importantly, all these solutions are written in human-readable code. Engineers can read, understand, debug, and predict algorithm behavior, which facilitates deployment to production.

What This Means

AI agents are moving from research laboratories into industrial engineering. AlphaEvolve shows that LLMs can participate in fundamental architectural decisions that require deep understanding of computer science. For developers, this means that part of the work to find optimal algorithms can be delegated to AI, while people focus on higher-level tasks.