Topic

From Prompt to Deliverable: Building an Agentic Pipeline for Batch Generation of Playable Ads

Playable ads deliver outstanding performance, yet they have long been constrained by the productivity bottleneck of being “handcrafted one by one.” The core contradiction is not that the models are not powerful enough, but that the non-deterministic outputs of large models cannot be directly integrated into production pipelines that require deterministic outputs. This presentation will provide a complete breakdown of a deployed Agentic generation system: how we use a Harness architecture to constrain the freedom of models, how we leverage a closed-loop sandbox + two-layer automated review to give the pipeline self-healing capabilities and achieve unattended content generation, and how we use a modular “Game Cartridge” specification to turn generated outputs into standardized deliverables that can be distributed to multiple channels with one click. The content includes real architecture diagrams, reviews of key failure cases, and transferable engineering paradigms. It will not only discuss “what AI can do,” but rather “how to make it do the right thing stably and at scale.” Outline Industry Pain Points and Engineering Challenges: Why are playable ads “amazing in performance but difficult to scale”? Harness Engineering Architecture Design: Building a deterministic engineering “saddle” for uncertain models Closed-Loop Sandbox and Two-Layer Automated Review: Achieving unattended self-healing and quality control Modular Delivery and One-Click Distribution: The “Standardized Game Cartridge” architecture Industry Implementation and Future Outlook Key Takeaways Master the engineering methodology for moving from standalone AI to industrial-grade Harness: Gain an in-depth understanding of the underlying architectural thinking behind moving from “prompt engineering” to “Harness guardrail engineering,” and learn how to use deterministic engineering code to tame the uncertainty of large models, avoiding the predicament of being able to produce only “toy demos” that cannot be deployed at scale. Gain an end-to-end implementation solution for a multi-agent automated pipeline: Gain a comprehensive understanding of how planning, art, coding, and review Agents are decoupled and coordinated through strongly typed Schema contracts, and learn how to effectively address long-context expansion and Agent role boundary violations. Learn the core design of sandbox self-healing and multimodal two-layer quality inspection: Gain an in-depth understanding of how to use a real headless browser sandbox to capture runtime errors and trigger targeted retries (Targeted Re-prompting), as well as how to build a two-layer quality review system combining multimodal LLMs with runtime metrics. Gain practical insights into standardized distributable cartridges and a gameplay decoupling architecture: Learn the asset packaging standards and implementation experience of the “Standardized Game Cartridge” (Game Cartridge), understand how to decouple configuration from the underlying engine, and enable one pipeline to rapidly derive multiple categories of mini-games and distribute them to various advertising platforms with one click.