Video director agents orchestrate multi‑stage production to realize users' creative visions. Recent work shows that self‑evolution can markedly improve screenplay planning, yet the mutual reinforcement between directing and user interaction is largely ignored. Better interaction clarifies user intent and guides workflow refinement, while more reliable execution shifts collaboration from troubleshooting to creative decision‑making. To address this gap we introduce LIT (Learning to Interact and DirecT), a framework composed of a director agent, a simulated user, and a meta engineer. The simulated user assigns creative tasks to expose the director's weaknesses; the meta engineer uses the task traces to revise the director's interaction and production strategies. Although this approach can fix discovered issues, it still relies on passive attribution by the meta engineer, limiting its adaptability to broader interaction patterns and production techniques. To overcome this limitation, LIT adopts a co‑evolution paradigm, enabling the meta engineer to proactively acquire external knowledge, thereby improving its ability to detect problems and propose solutions. The meta engineer, simulated user, and director form a closed‑loop evolutionary system that jointly enhances directing and interaction capabilities. Experiments on Test and OOD Test splits show that LIT improves production scores by 19.1% and interaction scores by 32.0% over the initial pipeline. Controlled ablations reveal a positive correlation between production and interaction improvements, and the co‑evolution mechanism further encourages exploration of directing skills and interaction modes.
Review: LIT demonstrates that empowering the meta engineer to learn actively enables the director system to self‑optimize, highlighting the promise of jointly evolving interaction and execution for future multimodal creative platforms.