A year after its debut, AlphaEvolve—a Gemini‑powered coding agent built by Google DeepMind—has moved from research demo to production work across Google’s stack and external partners. The system designs and refines algorithms for everything from DNA‑sequencing models to power‑grid optimizers, proving that automated algorithm discovery can replace months‑long engineering cycles with days of AI‑driven iteration.

In health research, AlphaEvolve cut variant‑detection errors in the DeepConsensus model by 30%, letting PacBio’s scientists extract cleaner genomic signals at lower cost. Grid‑management saw feasible solutions to the AC Optimal Power Flow problem jump from 14% to over 88%, slashing post‑processing overhead. Earth‑AI workloads gained a 5% boost in multi‑hazard risk forecasts after the tool streamlined geospatial optimizations.

Beyond science, AlphaEvolve now powers commercial pipelines. Klarna doubled transformer training speed, while Substrate accelerated lithography simulations by several folds. FM Logistic trimmed routing distances by 10.4%, saving 15,000 km yearly, and Schrödinger achieved roughly a 4× speedup in machine‑learned force‑field training. These deployments show the agent’s ability to translate algorithmic gains into measurable business value across sectors.

Internally, AlphaEvolve reshaped Google’s hardware pipeline, proposing a counterintuitive circuit that entered the silicon of next‑gen TPUs and refining Spanner’s LSM‑tree compaction to cut write amplification by 20%. Compiler tweaks derived from its suggestions trimmed software storage footprints nearly 9%, demonstrating how self‑optimizing code can tighten both performance and resource efficiency at scale.