RoboCoders: Judgment Day — AI Coding Agents Face Off

Abstract

Both agents can write code that runs. Neither can write code that is RIGHT — until you engineer the context they both inherit. And when you delegate, you have to engineer the context PROPAGATION too — because by default, sub-agents get nothing. Baruch brings Claude Code. Viktor brings GitHub Copilot in VS Code. Same tasks, same hardware (a smart bulb, a camera, two LED light bars), different agents. Five stages, escalating complexity: turn on a bulb, detect a face, recognize WHO, show confidence on a light bar, detect emotion via sub-agents. Both agents succeed at the easy stages. Both fail at Stage 3 — silently. The API returns 200, the code looks correct, but three segments on the light bar don’t physically exist. The confidence formula is textbook-correct and demo-wrong. The fill direction is upside down. The fix: four Tessl plugins encoding device ground truth, calibration data, actuator patterns, and camera setup — installed in one command. Same agent, same prompt, different context, four bugs gone. Then Stage 4: sub-agents lose everything. The concept of “emotion on the second bar” collapses because the IoT sub-agent doesn’t know what bar mapping IS. One meta-plugin that teaches the orchestrator explicit skill handoff restores the pipeline. This isn’t a framework comparison. It’s a live demonstration that the future of AI-assisted development isn’t about which agent is smartest — it’s about the engineering discipline around the agent.