An AI engineer that never stops experimenting.

CodeEvolver proposes changes to your prompts and code, tests each one against your data, and keeps only what wins — then proves it on data the engine never saw, and hands you a PR to review.

Four phases. One compounding loop.

01
Define Objectives
Set quality metrics, benchmarks, and success criteria. CodeEvolver uses evals as its north star — no vibes, only measurable signal.
eval.add("response_quality", fn=score_fn)
eval.add("latency_p95", threshold=200)
02
Curate Data & Evals
We build evaluation datasets from your observability logs, simulations, and domain experts — with ground-truth labels where the reward function needs them. No evals yet? That's part of the engagement.
03
Propose & Test
The engine generates candidate improvements — prompt revisions, logic changes, configuration updates — and runs them against your eval suite.
> Iteration 7 — Proposing change...
Score: 0.73 → 0.81 ↑ Accept
04
Validate & Deliver
When improvements plateau, winners are re-tested on held-out data to rule out overfitting. You receive a pull request with the full diff, score trajectory, and experiment history — your team reviews and merges.
> Holdout test: 0.78 confirmed
> PR opened: cev/phantomwiki-improvements

Software that rewrites itself — guided by evidence, not intuition.

A reflection agent analyzes failures in your system. A coding agent proposes candidate mutations to prompts, code, and model configuration. An evolutionary optimizer scores every candidate against your evaluation dataset and keeps only the winners.

The output is a pull request with the full experiment history and associated scores attached.

100s
Experiments per run
+80%
PhantomWiki relative gain
+85%
HoVer relative gain

What an actual run looks like.

A real optimization run on the PhantomWiki benchmark — three rounds, nine candidates evaluated, a 59% relative lift over baseline, validated on a holdout test set.

PhantomWiki Benchmark
CodeEvolver's Artificial Selection Algorithm explored and measured new ideas to improve system accuracy by 80%.
0.38+0% over baseline
0.350.450.550.650.75StartRound 1Round 2Round 3baselinechain-of-thought 0.41direct passage extract 0.52remove summarization 0.36rerank docs 0.60expand context 0.55swap base model 0.51factoid refinement 0.64structured output 0.62add reflection 0.69
codeevolver — optimization run
Start With a Pilot

See it run on your workflow.

Get a Baseline AssessmentSee the Research