Built with AIDCrew · 01 · supervised session
From a brief to a shared world.
Skyforge Islands is the browser prototype used to exercise AIDCrew's agent coordination, shared memory, Chromium plugin and synchronized terminal/web interfaces.
90 seconds of selected footage with English narration and a CC0 soundtrack added after the session. Audio credits. Build excerpts run at 20× speed; the final panorama includes a labelled build replay. The HTTP command overlay reproduces an actual submitted request with credentials omitted.
The team and the cost.
- Architect
- meta/muse-spark-1.3
$3.0372 - Coder
- z-ai/glm-5.3-flash
$1.2925 - Reviewer
- deepseek/deepseek-v4.1-flash
$0.6220
$4.95 in recorded agent API spend. All three models ran through OpenRouter. Shared memory was enabled, with one coder and milestone reviews. The TUI and Web UI displayed the same session.
Explore the per-role breakdown and measurement details in the session report.
A shared world. Verified progress.
Six automated tests passed on reviewed snapshots. Isolated two-client probes confirmed edit synchronization; same-database restarts preserved edits. A final normal-controls preview tour was recorded.
Trade-offs, limitations & lessons learned
The house was left unfinished: incorrect materials, missing walls and no roof. Spawn selection and documentation also had gaps. The user cancelled construction and closed the session on the available prototype.
The final agent commit is aba6e52; earlier review evidence is not a full independent review of that commit. Real-device touch controls remain unverified. Movement is client-authoritative.
Where the extra effort went.
At the first sampled checkpoint for the house phase, the session had spent $4.09. It closed at $4.95: an increase of about $0.86, including construction attempts, input fixes, coordination, review and the closing tour. The house was already in the brief. This phase total cannot be attributed solely to model errors.
The coder repeatedly rotated the camera while attempting a placement rejected as overlapping the player. Changing camera values hid the repeated outcome from exact-result checks. The review caught the difference between the coder's claim and the saved world. AIDCrew 0.3.0 adds a bounded stop for repeated page-reported rejection outcomes and gives queued guidance priority over automatic continuations. These fixes follow the recorded run; the footage does not claim to demonstrate them.
Supervision and reproducibility.
This session included 21 additional messages: 16 to the architect and 5 to the coder. Four interruption episodes included six cancel requests. These counts overlap and should not be added together. Instructions covered clarification, corrections, unblocking and the final cancellation/tour.
The recording spans about 2 hours 34 minutes, including idle periods. A full active-harness-time estimate is not asserted: the monitor began after the run started. The edited video's duration is not a development-time claim.
The source export starts a fresh world. It excludes session databases, private recovery QR codes, credentials, traces and raw recordings. Packaging adds license notices and corrects documentation; the game implementation remains the recorded agent snapshot.
git clone https://github.com/antoniociccia/aidcrew
cd aidcrew/demos/aidcrew-blocks/source
bun install
bun run devOpen localhost:8787. For deployment beyond a local demonstration, review the documented authentication, movement and input limitations first.