An optimization demo can make the next step look obvious: point an AI system at a slow function and wait for a faster one. That is not the workflow Google Cloud described when it made AlphaEvolve generally available on the Gemini Enterprise Agent Platform in July 2026. Its define–measure–optimize–apply loop starts with a seed program and a deterministic evaluator that runs on the client side.
That evaluator is the real admission ticket. “Faster without breaking anything” is not executable. A search process needs to know what it may change, what score it should improve, and which outcomes are disqualified even when the score rises. InfoQ’s independent report makes the boundary explicit: this approach fits problems whose success can be measured and automated. It also notes that vendor results are not independent benchmarks and that a search can exploit anything the evaluator forgets to test.
Write the contract before you spend the search budget
A useful evaluator contract has five jobs. It freezes a baseline and environment, computes one objective metric, enforces non-negotiable constraints, stops after a stated budget or plateau, and names the person who will inspect the surviving change. These are not administrative fields. Together, they translate “better” into a result the machine can test without pretending that the machine owns the release.
Klarna Engineering’s account shows why the omissions matter. Its engineers bounded the mutable code, metrics, and constraints. When determinism was left out, the search found a fast route that could not be deployed. The team also recorded 631 evaluations with no improvement, a practical reason to encode a plateau rule rather than letting the run continue until someone loses patience.
Route the target, not the tool
- Can the result be scored and hard constraints checked automatically?
- Objective score; all hard constraints automatedReady—fixed scope and budget, human review, no auto-deploy
- Score works; acceptance remains tacitNeeds evaluator work—encode one rejection reason
- Judgment is ambiguous or failure cost uncontainedUnsuitable—AI explores; people decide and own
This routing question comes before model choice. If no one can explain and own a generated patch, use the ownership price check for AI code changes first. If the evaluator is sound but the run still lacks deliberate pauses, add the human checkpoints required for coding agents to the operating plan.
A passing score is evidence, not release authority
Keep the seed version, environment, scores, constraint results, stopping reason, and final diff for every run. The named owner reviews that record and decides whether the candidate deserves broader security, operational, and product checks. An evaluator can reject code that violates encoded rules; it cannot prove that the rules captured every requirement.
Start smaller than a platform rollout. Pick one bounded optimization target and draft its baseline, metric, hard constraints, stopping rule, and human owner. If one field remains vague, the next task is to improve that field—not to generate more candidates.
AI handoff card
Begin with read-only discovery in the workspace, repository, test results, profiling output, or project records you can currently access. Without changing code, configuration, or deployment state, identify one bounded optimization opportunity and quote the direct evidence that led you to it; label observations separately from inference. Draft an evaluator contract covering the current baseline, one computable metric, correctness/determinism/resource constraints, a budget or no-improvement stopping rule, and the person who owns review and release. Mark missing facts “To confirm.” If essential context is inaccessible, ask no more than one targeted access or context question instead of requesting a prepared input package. Classify the target as “Ready,” “Needs evaluator work,” or “Unsuitable for automated search,” then propose one reversible next step that can be completed now. Do not deploy or publish a candidate, and do not treat a passing score as approval; the named human must inspect the diff and constraint evidence first.
Bound the search before judging the result
- The gardener begins with a small, manageable patch and compares only a few seedlings.
- Rows spread beyond the original trial area, and he stops to assess the uncontrolled expansion.
- He uses stakes and rope to define the single plot that the experiment is allowed to change.
- One bounded plot is healthy and complete; the rest of the field stays untouched for a later decision.
References
Google Cloud: AlphaEvolve is available for everyone — https://cloud.google.com/blog/products/ai-machine-learning/alphaevolve-is-available-for-everyone (2026-07-09)
InfoQ: Google’s AlphaEvolve Reaches General Availability with Evolutionary Code Optimization as a Service — https://www.infoq.com/news/2026/07/alphaevolve-generally-available/ (2026-07-19)
Klarna Engineering: Beyond Prompting: How Algorithmic Evolution Doubled our Training Speed — https://medium.com/klarna-engineering/beyond-prompting-how-algorithmic-evolution-doubled-our-training-speed-8f874af3080d (2026-03-30)
