Built a graph-based agentic video system that works like a real editor

Grapple is a graph-based agentic video platform developed solo over six months by a university student during a gap year. The system uses multiple Gemini 3 Flash agents—including planning, writing, audio for voiceover, video, and audio for music—managed by a main orchestrator agent called Steward. Unlike traditional pipelines, Grapple is stateful: it maintains a structured representation of the video, tracking exactly what changed between edits. When a user modifies the script, the system understands the change, triggering a ripple effect. The voiceover agent updates the audio, which then adjusts timing, which in turn updates video cuts. Each change propagates naturally through the graph, one step at a time. The name "Grapple" combines "graph" and "ripple."

Agents only access relevant nodes, avoiding dumping the entire video into context, which reduces tokens and latency. Users can prompt for an initial draft, then refine iteratively. For surgical edits, they can use commands like /audio or /video to modify specific components without affecting the rest. The workspace is shared in real-time: agents' changes (e.g., moving a clip) appear instantly, and user edits are visible to agents immediately. The creator notes that the main bottleneck is LLM taste—models tend to produce technically correct but editorially flat videos, as they are constraint-satisfying machines. Tightening constraints improves quality but reduces generality, a challenge the creator is still exploring.