Ray Fernando tests Gemini 4 Argon as an orchestrator while keeping Opus for delegated coding work. He adjusts model roles, review panels and effort levels in a live development session, rather than presenting a controlled model ranking. His central question is how different models can work together across planning, implementation and verification.
Ray Fernando develops an experimental robotics simulation with a proposed local vision-inference loop. He discusses a small robot, synthetic care environments and possible notifications from detected falls. This is software experimentation and planning, not a demonstrated physical care robot or a validated medical, privacy or emergency-response system.
Ray Fernando explains skill routing, reusable task instructions and multi-model review, while reducing parallel review roles to control cost. He argues that short coding or game benchmarks miss the coordination, context and dynamic environments of longer agent workflows. Those observations are his interpretation of selected experiments, not independently established comparative performance.
Ray Fernando also requests a native voice-input app connected to his agent setup and plans a larger software refactor. The app initially heads toward a web interface and later produces an awkward, intrusive interface that he repeatedly corrects. Voice handoffs begin working during the session, but the application still needs design direction and the separate refactor remains unfinished.
Watch on YouTube
