Video Generation Is Just the Beginning: How Can AI Get Content Right

Originally published by InfoQ

Creating a video used to mean learning editing software, choosing a template, adjusting shots and adding finishing touches step by step. These tools lowered the barrier to video creation, but they did not answer a more fundamental question: What is the video actually meant to achieve?

On October 22, at the main forum of QCon Shanghai 2026, Han Sheng, Founder and CEO of QuVideo, shared the company's long-term thinking on content creation in a keynote titled "From Tools to Agents: Three Paradigm Shifts in QuVideo's Content Creation Journey".

In his view, AI is changing content creation beyond helping users generate individual assets. It is moving toward understanding users' goals and delivering finished results.

1. From Editing Tools to Content Agents That Deliver Results

Founded 14 years ago, QuVideo has long served the video creation needs of users around the world. Han noted that QuVideo's products have accumulated more than 2 billion downloads. As user needs and technology have evolved, the company's content creation products have gone through three stages.

The first stage: video editing tools. From 2012 to 2022, QuVideo primarily helped users edit videos through mobile applications. Users remained in control of the creative process, while the products provided features such as timelines, music, transitions, filters and subtitles to make video production easier.

The second stage: AIGC applications. Starting in 2022, generative AI became part of the content production process. Users could generate assets and content more quickly through capabilities such as text-to-image, image-to-video and image-to-music generation. Models began assisting with some of the work traditionally done by designers, editors and professional creators.

However, Han sees AIGC applications as largely a new generation of more powerful tools. Users still need to come up with ideas, choose the right features, organize the workflow and take responsibility for the final result.

The third stage: content agents. Since 2025, QuVideo has been exploring how AI can move beyond generating individual assets to understanding goals, using tools and delivering complete results.

The product's central focus shifts from features to tasks. When a user says, "Create a promotional video for me," the agent needs to understand the product, target audience, brand, distribution channel and marketing objective. It can then select appropriate models, assets and workflows to deliver a video that is ready to use.

The central question for content creation products also changes: instead of asking how to help users use tools more effectively, the focus becomes how to help them get something done.

2. Why Is Generating a Video from a Single Prompt Still So Difficult?

Models are becoming increasingly capable, but the idea that ordinary users will soon be able to generate any video they want simply by describing it in one sentence remains difficult to realize. Han identified three key challenges for video creation agents: open-ended user instructions, the difficulty of automatically evaluating results, and maintaining consistency across long contexts.

First, ordinary users often cannot fully articulate what they need. Han repeatedly emphasized that what users say is often only a surface-level expression of their actual needs. A request such as "Create a marketing video for me" can represent very different objectives.

A nail salon owner may want to attract customers and drive bookings. A bubble tea shop may care more about promoting discounts and encouraging immediate purchases. An e-commerce merchant may want to communicate a product's selling points clearly, while a content creator may be more concerned with keeping viewers engaged. Although all these requests sound like "making a video," they involve different Jobs to Be Done, creative workflows and measures of success.

Second, video results are difficult to evaluate against a universal set of rules. Software development tasks can often be assessed through test cases and execution results. Video creation, however, has no single standard answer that applies to every scenario. Visual quality, messaging, brand consistency, whether users accept and publish the video, and even whether it leads to bookings or transactions may all serve as evaluation criteria.

This means an agent must do more than generate a video. It must understand what a successful result looks like in a specific context. General-purpose models alone cannot fully define that standard. Companies also need to turn their accumulated creative experience, scenario knowledge and business expertise into executable, reusable evaluation criteria.

Third, video creation requires contextual consistency across shots, assets and stages of production. People, settings, products, prices, phone numbers and brand information all need to remain consistent across multiple shots. Users may repeatedly change a price or phone number during the creative process. The agent must understand the intent behind these changes and ensure they are reflected throughout the final video.

For these reasons, QuVideo has not simply set out to make models generate more content. Instead, it starts with real-world scenarios where results can be evaluated, then turns validated workflows and professional expertise into reusable Skills.

For ordinary users, the key experience is not learning a new set of prompts. It is being able to say: "This is exactly what I wanted to create, and I can get it done quickly."

Recent data on consumer AI adoption also supports this view. According to a 2026 consumer AI report from a16z, the range of AI use cases among U.S. consumers is expanding, while deeper engagement and paid usage remain concentrated among a smaller group of high-value users and professional scenarios. Creative, coding and productivity tools are among the main categories for which users are willing to pay.

This suggests that consumer AI products need clearer value propositions than general-purpose chat. For QuVideo, the priority at this stage is to start with specific content creation and business scenarios, using productized Agents to address concrete needs in users' day-to-day work.

3. From Model Integration to AgentOps

As products evolve from individual tools into multiple Content Agents, the challenge is no longer simply connecting to a more powerful model. It is ensuring that different business applications can operate reliably and compliantly.

QuVideo serves users across global markets, with different cultural backgrounds and creative habits. This brings several engineering challenges:

• Different regions have access to different models, resulting in variations in model performance, pricing and reliability.

• Specialized scenarios may require proprietary algorithms or LoRA models.

• Existing products need to continuously incorporate Agent capabilities.

• Content safety, privacy protection and platform compliance must be addressed before products go live.

Rather than having each business line build its own infrastructure, QuVideo is developing a unified AI Agent infrastructure platform: AgentOps.

Through a unified multimodal model gateway and intelligent routing, the platform brings proprietary models, third-party models and domain-specific models developed by business teams into a single system, reducing the cost of model integration and management. Within a shared engineering framework, business teams can select models and invocation strategies based on specific tasks, allowing them to focus more on task objectives and the quality of final deliverables.

Beyond model integration, capabilities such as Agent Runtime, visual orchestration, Skills and Tools, cost management, security governance, evaluation and observability are also organized into reusable components. New business applications can combine these components as needed, lowering the engineering barriers to building, debugging and launching Agents. This also turns algorithms and product expertise previously scattered across different teams into reusable capabilities.

For QuVideo, the value of AgentOps ultimately lies in the delivery process. The system continuously records and tracks Agent execution, including conversations, model calls, token consumption, generated results, errors and user feedback. This information is retained for issue diagnosis, performance evaluation and subsequent product improvements. It helps more Agents reach real-world scenarios faster and become more reliable, controllable and aligned with users' actual needs through continuous feedback.

4. From Sequential Development to Joint Product and Engineering Ownership

Traditional software development typically follows a sequential process: product definition, technical planning, coding, testing, optimization and release. AI can improve the efficiency of individual roles, but if the overall workflow remains unchanged, teams still have to wait for upstream stages to finish. End-to-end delivery speed will not fundamentally change.

QuVideo is exploring a different way of working.

Under the AgentOps model, product and engineering teams work together to build Agents around real user needs and specific business scenarios. Product teams stay close to users and business operations, identifying needs, defining task objectives, and participating in Agent assembly and scenario validation. Engineering teams provide the underlying support, including model capabilities, Agent Runtime, model gateways, evaluation platforms and security mechanisms. Both teams work together to develop reusable Skills, improve performance and validate quality.

This approach allows product ideas to be tested quickly in real-world scenarios, while engineering capabilities are continuously accumulated and reused across different Agents.

One example is Kiva, an Agent that QuVideo has already incorporated into its delivery workflows. After users provide text, images and voice samples, the system organizes the content around specific business objectives to produce a video.

Once a product enters real-world use, the core value of AgentOps goes beyond coordinating models and tools. It also lies in turning user behavior into reusable scenario assets: identifying who the user is, what task they want to complete, what their key needs are, what constitutes success and how the scenario should be prioritized. These structured scenario insights can then serve the testing, product and marketing teams.

For example, a user may repeatedly change a phone number or price while creating a video. These seemingly minor actions can reveal important signals about the user's needs:

• Product team: Identify the user's specific scenario and trace whether the issue lies in the interaction, workflow or content generation process.

• Marketing team: Use business information and creative scenarios voluntarily provided by users to identify potential business needs and analyze common patterns across similar scenarios.

• Testing team: Evaluate different product versions against the same task to determine whether the product has actually addressed the user's needs.

These scenario insights then enter the evaluation process. Once a video has been generated, the system must answer another question: Does the content actually meet the requirements?

Meanwhile, the team builds datasets for different scenarios. Evaluation agents test multiple versions in batches, followed by human review and final calibration. The results are then used to refine prompts, Skills, tools and model selection, creating a continuous cycle of generation, evaluation and optimization.

In the past, understanding what users liked often meant manually reviewing examples and relying on experience to draw conclusions. Now, the system can regularly summarize usage patterns, user volumes and actual results across different scenarios, making it easier to see which scenarios are gaining traction, which needs occur most frequently and where problems remain.

Teams can then assess the reach, frequency and performance gaps of different scenarios. High-frequency scenarios are prioritized for refinement and turned into reusable Skills, while lower-frequency scenarios can be covered through guided experiences and templates. User behavior thus informs product optimization, market insights and performance evaluation at the same time, providing a shared basis for continuous agent improvement.

This also changes how teams collaborate internally. Engineering teams no longer need to repeatedly develop individual features for separate business lines. Instead, they build shared platforms and components that support multiple products. Product teams, meanwhile, do more than submit requirements: they participate in model selection, agent assembly, performance validation and scenario development, sharing responsibility with engineering for the final results.

5. User Behavior as a New Product Asset

In QuVideo's experience, important signals for an agent do not come only from what users say. They also come from how users make changes, repeat actions and decide whether to use the results during the creative process.

QuVideo organizes these signals into user profiles, scenario tags and Jobs to Be Done (JTBD) insights, and makes them available to testing, product and marketing teams. Testing teams can reproduce issues based on the same scenarios; product teams can identify specific stages for optimization; and marketing teams can reach audiences based on clearly defined needs.

In this way, a real creative session no longer ends with a single deliverable. It becomes part of a continuous cycle: real-world usage → behavior analysis → scenario insights → Skill release → performance validation. Product assets now extend beyond features and code to include user preferences, success criteria and expert knowledge.

There are also clear boundaries around how this information is used. Only behavioral information for which users have given explicit consent enters this cycle for product evaluation and continuous improvement. Original creative materials are strictly classified and managed according to user authorization, data access permissions and the compliance requirements of individual markets. Respecting user boundaries is essential to making user behavior a legitimate and sustainable product asset.

In Han's view, this closed loop ultimately serves two goals: validating needs faster and ensuring that each iteration leaves behind reusable knowledge. He summarized QuVideo's next-stage priorities in two points:

"First, rapid iteration. Second, cumulative learning."

As scenario data, user preferences and evaluation criteria continue to accumulate, these assets can be translated into stronger product capabilities, helping QuVideo better serve users across different countries and markets.

6. As Models Become General-Purpose, Scenario Expertise Creates Differentiation

As general-purpose models continue to evolve, model capabilities themselves are increasingly becoming infrastructure for AI applications. During his talk, Han observed that general-purpose models will eventually become like "utilities" - as ubiquitous as water, electricity and gas.

Against this backdrop, QuVideo is pursuing a more focused approach: turning its 14 years of experience with global users, content preferences across markets, real-world creative scenarios, user behavior data and delivery standards into lasting company assets.

Models provide general-purpose capabilities. Agents understand tasks and coordinate execution. Platforms ensure reliability. Data and Skills preserve and reuse scenario-specific knowledge, enabling systems to better understand users through continued use.

The impact of content agents therefore extends beyond video products themselves. They are also driving changes in product design, technical architecture, evaluation systems and R&D collaboration.

In the past, users had to learn how to use tools, while teams built products around individual features. Now, users express their goals directly, agents take on the tasks, and teams continue to build capabilities around scenarios, platforms and data.

During his talk, Han also raised a fundamental question: Where does the competitive advantage of an AI application company really lie? If a product simply wraps a general-purpose model, its capabilities can be easily replicated. What is much harder to replace is a company's accumulated understanding of its users, scenarios and standards for success.

This is the core shift behind QuVideo's move from tools to agents: competition in content creation is moving from the number of features a product offers to the results it delivers. Companies that better understand what users want to accomplish are more likely to turn AI into a reliable, reusable product capability.

What companies retain is the understanding of users, delivery experience and continuous improvement capabilities built through real-world scenarios.