Rexgent, a project published on Alibaba Cloud's community blog, is an autonomous short-drama production pipeline built as a LangGraph graph of agents over 17 Qwen Cloud models. A user supplies a premise or a screenplay, picks one of 22 visual styles, sets a budget and an episode count, and the system carries the production through script, character design, shot planning, rendering and assembly. According to the write-up, the pipeline targets continuity rather than raw generation quality. Character identity plates are generated from one reference photo to lock faces and outfits across shots; an AI director breaks scenes into shots with shot size, camera angle, lens and blocking while enforcing the 180-degree rule; a Neo4j narrative memory is used to keep later scenes from contradicting earlier ones; each rendered clip continues from the previous frame and is scored for drift, with flagged takes sent for review. Model choice is routed by task: Qwen-Max for story development and direction, Qwen-Plus for analysis such as narrative judging and relationship extraction, Qwen-Flash for high-volume parsing and extraction, Qwen3-VL-Plus for clip continuity scoring, Wan2.6-T2I and Qwen-Image-Edit-Max for images, and separate video models for character shots, scenery and repairs. The author reports building it alone on 40 dollars of Qwen Cloud credit and argues the budget forced planning ahead of rendering: script scoring, face locking and shot planning all happen before any video is generated.
Our takeThe transferable idea is the cost ordering: cheap text and vision passes go before expensive video generation, and continuity is enforced with artifacts, namely identity plates, frame handoff and drift scores, rather than by asking a model to remember. That is what stops a multi-stage generation pipeline from quietly rewriting its own characters. What is missing is measurement: no per-stage failure rates, no comparison against a human editor, and the 40-dollar figure describes a budget, not a quality bar.