612 Commits in 24 Days.
What did we learn?
A case study on what an agent-driven engine port taught us about AI, verification, and production workflows.
Inside, you'll find:
- How Amber structured a five-role AI agent harness around one senior engineer
- Why verification against ground truth mattered more than raw code output
- How the team tested data, logic, behavior, and live-play parity
- What broke despite “green” test gates, and how the workflow changed in response
- What the four-week velocity curve revealed about scaling agent-driven development
- The production principles Amber is taking forward from the experiment
Over 24 days, one senior Amber engineer directed AI coding agents to rebuild the opening sequence of The Elder Scrolls III: Morrowind in Godot, using the open-source OpenMW codebase as the reference. The result was a playable end-to-end slice, backed by a growing verification system designed to test the port against the original rather than against the agents' own assumptions.
The experiment quickly became about more than speed. As the team moved from a single agent and chat window to a multi-agent pipeline, the real work shifted toward orchestration, ground-truth verification, adversarial review, persistent project memory, and keeping humans in the loop where automated gates were not enough.
By the end of the four weeks, the project had produced 612 commits and a pipeline Amber plans to carry into production work: plan, delegate, verify against a reference, and review before merge. The experiment was deliberately narrow, and the team documents the remaining gaps as openly as the progress.