AlphaEvolve reaches GA as a code-optimization agent that wants an evaluator, not a prompt
Google moved AlphaEvolve, its Gemini-powered algorithm-discovery agent, to general availability on July 9. Unlike a chat coding assistant, it optimizes against a scoring function you supply: powerful for measurable problems, inert without a good evaluator.
Google moved AlphaEvolve to general availability on July 9, making its Gemini-powered algorithm-discovery agent something any team on the Gemini Enterprise Agent Platform can run, rather than a research demo or a private preview. The framing matters more than the milestone: AlphaEvolve does not answer prompts. It takes a working algorithm and a way to score it, then evolves faster or cheaper versions against that metric. That single requirement, a real evaluator, is what separates the teams it will help from the ones it will frustrate.
What shipped
AlphaEvolve is now GA as a “code optimization and discovery agent built on top of Gemini,” aimed at the class of problems where a better algorithm is worth real money: routing, forecasting, kernel and compiler tuning, chip design, scheduling. The workflow is a four-step loop. You define a seed algorithm and the problem, you supply a scoring function that measures correctness and performance, AlphaEvolve’s agentic harness generates and mutates candidate code, and you deploy whatever beats your baseline. DeepMind first introduced the system in May 2025; GA is the point where it stops being a Google-internal and preview capability and becomes a product teams can buy.
Access is broader than the enterprise console suggests. Beyond the Gemini Enterprise Agent Platform, there is an AlphaEvolve API, a public GitHub repository with runnable examples that range from local circle-packing to a GKE-and-Terraform fine-tuning setup, and an AlphaEvolve Skill that plugs into IDEs including Google’s Antigravity and, notably, Anthropic’s Claude Code. Shipping a skill into a competitor’s agent is a deliberate reach for developers wherever they already work.
The results, and the catch
The customer numbers Google cites are the reason to pay attention. JetBrains reports 15 to 20 percent IDE performance gains; Kinaxis claims a 22 percent forecasting-accuracy improvement and a 90 percent runtime reduction on benchmark datasets; FM Logistic trimmed warehouse routing by 10.4 percent, saving more than 15,000 km of staff travel; Klarna doubled throughput in ML training pipelines. Internally, Google says AlphaEvolve cut Spanner’s write amplification by 20 percent and reduced some software storage footprints by 9 percent. These echo the earlier research results that made AlphaEvolve’s name, including a matrix-multiplication algorithm that beat a 1969 benchmark for one case and a scheduling heuristic that recovers roughly 0.7 percent of Google’s worldwide compute.
The catch is in the setup. AlphaEvolve needs a deterministic evaluator, a client-side script that compiles, runs, and scores each candidate. That is straightforward for problems with a clean objective and expensive for problems without one. If you cannot express success as a number, AlphaEvolve has nothing to optimize toward. This is the opposite constraint from a general coding assistant, which will attempt anything and guarantee nothing. AlphaEvolve guarantees improvement against your metric, provided you can write the metric, and provided the metric cannot be gamed by code that scores well without being correct.
Where this lands
This is a different shape of coding agent than the ones most teams have adopted. GitHub Copilot, Cursor, and Claude Code optimize for developer throughput across open-ended work. AlphaEvolve optimizes a narrow, measurable target and leaves the rest alone. For the right problem, a hot inner loop, a routing heuristic, a training kernel, that focus is the point: it can search a space no human has time to explore and hand back code that provably scores higher. For the wrong problem, it is overhead with a setup cost.
Pricing is the missing variable. Google has not disclosed what AlphaEvolve costs at GA, and the value case depends on it. An agent that shaves a few percent off a workload is compelling if the run is cheap relative to the savings and marginal if it is not. Teams with a specific, costly, measurable bottleneck have the clearest reason to pilot it now; everyone else can reasonably wait for pricing and independent results.
What’s worth watching
- Pricing at GA. Undisclosed today, and central to whether AlphaEvolve is worth pointing at anything short of a very expensive problem.
- Whether teams can write good evaluators. The technology assumes a clean scoring function. How many real production problems can be expressed that way, without gaming the metric, is the practical limit on adoption.
- The cross-platform play. An AlphaEvolve Skill inside Claude Code puts Google’s optimization loop next to Anthropic’s agent. Whether that kind of interoperability spreads, or gets walled off, is a signal about how the agent market shares surfaces.
The plain read is that AlphaEvolve is not a Copilot rival; it is a specialist that turns measurable, expensive problems into a search you can automate. Stackmaven will revisit it once pricing is public and independent teams report results on their own workloads, on or around October 8.