Great project! have you considered giving the agents code-based video rendering like MoviePy or Remotion? Since LLMs are mostly trained on texts and code maybe for edits use code based approach and for complex scenes fallback for seedance, elevenlabs mcp or connectors