At Helix, the team initially developed a sophisticated H.264 streaming pipeline using WebSockets for real-time video sharing of AI coding assistants. They achieved impressive specs: 60 frames per second at 40 Mbps with under 100 ms latency. However, when deployed in enterprise environments, the constraints of corporate networks, which often block UDP traffic essential for WebRTC, led to significant challenges. The video stream performed well in ideal conditions but suffered delays and buffering when network conditions worsened, resulting in frustrating lag for users. In a moment of desperation, the team pivoted to using JPEG screenshots instead of video. They found that JPEGs, being self-contained, could deliver reliable images regardless of network quality. Each screenshot averaged about 150 KB, while H.264 keyframes could be 200-500 KB. This shift not only reduced bandwidth usage but also improved reliability. They implemented an adaptive system where, under good conditions, the high-quality video would run, while in poorer conditions, the system would switch to polling for JPEGs. Their method allowed keyboard and mouse inputs to remain responsive through the WebSocket, even when video was paused. However, they encountered a "oscillation problem" where the system would rapidly switch between video and screenshots. The solution involved locking the system in screenshot mode until the user manually requested a return to video. This approach struck a balance between maintaining quality video streaming and ensuring a smooth user experience, even in less-than-ideal network conditions.