Proposes a fully automated AI system using agent populations mimicking studio roles to generate SNL-style comedic videos via iterative competition, evaluation, and improvement. Key contribution: LLM critics aligned with real viewer preferences through preference analysis.
The concept of using AI agents to generate sketch comedy, such as SNL-style content, aligns with emerging developments in AI-driven creative systems. While no fully automated agentic system for sketch comedy is detailed in the provided context, AI is increasingly being used to generate comedic content through roles that mimic traditional production workflows. For example, ReelMind.ai features Nolan, described as the "World's First AI Agent Director," which functions as a virtual co-director by offering intelligent scene composition, narrative structure guidance, and cinematography suggestions . This mirrors a studio role and supports the idea of AI agents taking on specialized creative functions.
Additionally, platforms like ReelMind.ai utilize populations of AI models—over 101 in their library—including specialized video generation models such as Kling V2.1 Pro and PixVerse V4.5, which can create exaggerated movements and surreal visual gags essential for comedy . These models can be used in tandem with multi-image fusion to maintain character and prop consistency across scenes, a critical factor in recurring sketch formats .
AI systems are also being used to generate sitcom scripts by analyzing vast datasets of existing shows, with some implementations allowing human writers to use AI-generated content as a starting point for refinement, creating a hybrid human-AI collaborative model . One example is Botnik Studios, which uses predictive text technology trained on sitcom dialogue patterns to produce comedic sketches that blend absurdity and wordplay .
Although a fully autonomous system using competitive agent populations to iteratively generate, evaluate, and improve comedic sketches is not explicitly documented in the sources, the foundational components exist: AI agents performing directorial tasks , generating scripts , ensuring visual consistency , and enabling batch generation for rapid iteration . These capabilities suggest that an agentic framework for automated sketch comedy generation is a plausible extension of current AI tools.
COMIC: Agentic Sketch Comedy Generation introduces a fully automated framework for producing sketch comedy videos by simulating the collaborative and competitive dynamics of a professional studio environment. The paper details a multi-agent system where populations of LLMs adopt specific roles—such as writers, producers, and actors—to iteratively generate, critique, and refine scripts. Unlike traditional single-prompt generation methods, COMIC employs a structured workflow of competition and evaluation, allowing agents to pitch ideas, receive feedback, and improve content through multiple rounds. This process culminates in the generation of not only scripts but also the corresponding video assets, creating a complete end-to-end pipeline for SNL-style content.
The key technical contribution of this work is the development of LLM-based critics that are explicitly aligned with real viewer preferences through rigorous preference analysis. By training or tuning these critic agents to understand the nuances of human humor—specifically what makes a sketch successful or unsuccessful—the system ensures that the iterative improvements move toward higher quality, more entertaining outputs rather than simply optimizing for generic coherence or fluency. This alignment mechanism solves a persistent challenge in computational creativity: bridging the gap between grammatical correctness and subjective entertainment value.
This research matters significantly as it demonstrates the potential of agentic AI architectures to handle complex, multi-stage creative workflows that traditionally require human intuition and collaboration. It validates that distinct social roles and competitive structures can be effectively modeled within AI systems to produce sophisticated creative artifacts. For the field of generative AI, COMIC offers a blueprint for automating content creation in highly subjective domains, suggesting that future systems can move beyond simple text generation to manage the intricate production pipelines required for narrative-driven media.
This paper presents COMIC, a novel framework for fully automated sketch comedy generation that mimics the collaborative and competitive dynamics of a professional comedy studio (e.g., Saturday Night Live). The system deploys a population of AI agents, each specialized in distinct roles—such as writers, directors, and performers—to iteratively refine comedic content. The process leverages LLM-based critics to evaluate sketches, with these critics trained to align with human preferences through preference learning techniques (e.g., pairwise ranking). By simulating real-world creative workflows—including brainstorming, revision, and iterative improvement—COMIC aims to produce higher-quality, more engaging humor compared to traditional single-model generation approaches.
The key contribution lies in the agentic architecture, which decouples creative tasks (e.g., joke writing vs. delivery optimization) and introduces competition among agents to drive novelty and quality. Additionally, the use of preference-aligned critics ensures that the generated content reflects real audience tastes, addressing a critical limitation of prior AI comedy systems, which often produce laughs that feel contrived or misaligned with human humor. This work matters because it advances AI-generated entertainment by combining multi-agent collaboration with human-centric evaluation, offering a scalable model for automated humor production. The implications span content creation, digital media, and even AI-assisted creativity in broader applications.