Evolution Engine — autonomous product evolution

Showcase repository. The source code is private; this page documents the concept and design. Built by Roderick Specter.

Most AI coding workflows improve software by reading code and guessing at problems. Evolution Engine does the opposite: it uses the product like a real user, measures what happens, and changes code only when there is evidence that something blocks or degrades the user's goal.

---

The loop

flowchart LR
    G["1 · Observe a real user goal"] --> T["2 · Attempt the task<br/>in an isolated browser"]
    T --> S["3 · Gather multimodal evidence<br/>screens · DOM · console · network · timing"]
    S --> D{"4 · Does friction materially<br/>block or degrade success?"}
    D -- "No" --> N["NO_CHANGE_WARRANTED<br/>(a first-class success)"]
    D -- "Yes" --> F["Fix only that"]
    F --> R["5 · Re-run: baseline vs variant"]
    R --> P{"Better?"}
    P -- "Yes" --> PR["Promote"]
    P -- "No" --> RV["Revert"]

Loading

Core invariants

• Evidence before edits: no change without measured impact on a real flow
• No-op is success: if the product works, the engine changes nothing
• Experimental discipline: every change is an A/B comparison against a baseline, then promoted or reverted
• Deterministic bench: a built-in suite of healthy, broken and high-friction flows verifies the engine itself

Design

• TypeScript monorepo (pnpm workspaces) with a CLI
• Browser automation sandbox for task attempts
• Pluggable evidence collectors and evaluators

---

Reach me through my GitHub profile or Upwork.