At WAIC 2026, the large model and generative AI track had a total of 130 participating entities, with 83 of them focusing on AI Agent or intelligent applications as their main tags, leaving only 18 basic large model companies. The number of model releases has decreased, while agent demonstrations have increased. The focus of industry discussions has shifted from "how smart the model is" to "how much real value it can create." The fundamental unit of AI competition is evolving from "a single model" to "a system."

01       From Model to System: AGI Organizational Goals and Core Development Elements of AI

In 2026, the global artificial intelligence industry is standing at a delicate yet critical crossroads. Model capabilities continue to break through, intelligent agents accelerate their deployment in physical scenarios, and corporate AI strategies shift from scattered exploration to systematic integration. At the same time, discussions about AGI (Artificial General Intelligence) are shifting from "whether it can be achieved" to "how to achieve it"—and even more importantly, "which organizations can achieve it."

For current R&D organizations, this is no longer a matter of choice but a survival issue.

AGI as an Organizational Goal: From Vision Consensus to Strategic Anchor

AGI is not just a simple question-answering tool, but an intelligent system capable of understanding complex goals, decomposing tasks, calling tools, continuously learning, and performing high-level work in research, engineering, finance, manufacturing, and social governance. Global AI leaders such as OpenAI, Anthropic, and Google DeepMind all regard it as the ultimate goal of AI development.

However, the most significant change in 2026 lies in the fact that AGI is no longer just the "poetry and future" of a few cutting-edge laboratories—it is becoming the core strategic anchor for an increasing number of R&D organizations.

This July, Liang Wenfeng, founder of DeepSeek, systematically outlined the company's AGI roadmap in a four-hour closed-door exchange: a six-step progressive structure—language model → chain of thought (CoT) → Agent → continuous learning → self-iteration (singularity) → embodied intelligence. Liang Wenfeng emphasized, "Agents should use CoT, and CoT should also use previous steps, so none of these steps are wasted."

These signals reveal a consensus: AGI is not a sudden leap, but a technical evolution path that must be carefully advanced step by step.

More importantly, AGI is becoming the "first principle" of organizational decision-making. DeepSeek is a typical vision-driven organization—its operations, technology choices, and business decisions all revolve around the consensus of AGI R&D, relying on long-term goals to gather core R&D talent. Tang Jie also pointed out that the core essence of this AI revolution is not a product innovation or a business model innovation, but a technological revolution that has raised the "intelligent ceiling." Whoever can push this ceiling up by an inch first will redefine the capability boundaries of thousands of industries.

Core Development Elements of AI: From "Three Horsemen" to Systematic Computing Power

If AGI defines "where to go," computing power, data, algorithms, talent, and organizational structures collectively determine "how far one can go."

Computing Power: From Single Chip Competition to System Competition

Computing power is the foundational base and key engine supporting the continuous evolution of artificial intelligence. According to the Ministry of Industry and Information Technology, as of June 2026, China's intelligent computing power scale reached 2185 EFLOPS. However, the logic of computing power competition is undergoing a fundamental change.

Previously, people always asked about the level of a single chip. But large model training and inference are not single-chip competitions—interconnection, memory sharing, and task scheduling determine whether a large-scale cluster can form effective computing power. The focus of AI computing power demand is shifting from training to inference, and storage, advanced process nodes, and advanced packaging capacity are increasingly strained. This reveals: computing power provides the foundation, models provide intelligence, Agents handle execution, and terminals and robots enter real-world scenarios—the basic unit of AI competition is moving from "a single model" to "a system."

Data: From "Managing" to "Using"

Data is the key "fuel" for artificial intelligence. In 2025, the total amount of data used for AI training and inference in China was 199.48 exabytes, an increase of 42.86% year-on-year, with the volume of inference data surpassing that of training data for the first time. As of June this year, the total volume of high-quality industry datasets built nationwide exceeded 1565 petabytes.

But the value of data does not lie in "quantity," but in "usage." Liu Liehong, director of the National Data Administration, pointed out that the topic of data collaboration is moving from "managing data" to "using data." For R&D organizations, breaking through industrial chain data barriers and promoting data aggregation and shared governance is necessary to enable AI to accurately take root in industrial scenarios.

Algorithms: From Micro-innovation to Fundamental Breakthroughs

The field of algorithms has reached international advanced levels. But the real challenge lies in: the competitive focus of future large model companies should not merely be micro-innovations within existing technical frameworks, but rather fundamental breakthroughs in the cognitive ability, reasoning ability, and integration of world knowledge of the model.

In 2026, the industry consensus is shifting from language models to multimodal world models that understand physical laws. The ability to handle long-range tasks—enabling AI to move from "instant Q&A" to "grand projects," autonomously breaking down grand goals into thousands of executable subtasks—is a cutting-edge direction for algorithmic breakthroughs.

Talent: The Most Scarce Resource

No matter how powerful the computing power or how much data there is, it ultimately relies on people to manage. Data shows that the supply and demand ratio for high-performance computing engineers is as low as 0.15—each qualified job seeker faces competition from seven companies. AI scientists have an average monthly salary of 132,000 yuan, leading the pack by a wide margin.

But talent competition is no longer simply about "hiring people." Among top global AI researchers, 47% are Chinese nationals. How to retain talent and stimulate creativity is a test of the organization's systems and culture. DeepSeek's approach is to use the long-term goal of AGI to unite core R&D talent;

Organizational Structure: From "Racing" to "Unifying Forces."

One of the most noteworthy trends in 2026 is the systemic restructuring of AI R&D organizational structures.

Over the past two years, internet giants followed internal racing, with multiple AI product lines testing different approaches, consuming computing power and human resources continuously. However, by mid-2026, the big players collectively pressed the integration button—ByteDance merged the Feishu product team into Douyin, Alibaba integrated multiple Agent product lines and launched Qianwen Office, and Tencent consolidated its R&D resources.

This is not unique to China. Amazon shut down its AGI Lab research team, merging the AGI department, in-house chip team, and quantum computing team into a unified organization, integrating the entire supply chain from infrastructure to model development. Google DeepMind implemented an organizational structure adjustment, optimizing the division of research and operations, and concentrating resources to advance AGI technology development.

These adjustments reflect a profound judgment: the core of AI competition has shifted from single-point technological breakthroughs to foundational model capabilities. A leading model needs to simultaneously possess multi-dimensional capabilities including algorithms, data, computing power, training, engineering, and product feedback—this capability enhancement no longer relies on a single laboratory but on the entire organization forming a high-speed cycle.

From "Acquiring Intelligence" to "Organizational Intelligence"

The AI industry is moving from "acquiring intelligence" to "organizational intelligence." Models still determine the upper limit of AI capabilities, but relying solely on a leading model is increasingly difficult to constitute a complete industrial advantage.

For today's R&D organizations, AGI is not just a slogan, but a system that requires systematic construction—requiring organizations to evolve synchronously in computing power, data, algorithms, talent, and architecture, requiring decision-makers to have sufficient strategic perseverance to make choices between short-term profit and long-term breakthroughs, and also requiring every participant to understand that the path to AGI cannot be skipped.

As written on DeepSeek's recruitment poster—"humans are currently at the dawn of AGI. But the dawn is not the dawn." The organizations that can cross this long night are those who have clear goals and can solidly build core development elements as long-termists.

02       Why Agents Will Eventually Return to the Model Core, Pre-training as the Ceiling

In 2026, the clearest line of development in the AI industry is undoubtedly the comprehensive leap of Agents from "chatting" to "doing work." The overall level of base models has risen from 20 points to 70-80 points, and the stable execution time for long-range tasks doubles every eight months, with the highest record reaching 16 hours. Agents no longer "ask one question and answer one question," but instead complete the loop of "observation → thinking → action" independently at each step, allowing the large model to serve as a reasoning hub, autonomously breaking down tasks, calling tools, and evaluating results.

At WAIC 2026, the large model and generative AI track had a total of 130 participating entities, with 83 of them focusing on AI Agents or intelligent applications as their main tags, leaving only 18 basic large model companies. The number of model releases has decreased, while Agent demonstrations have increased. The focus of industry discussions has shifted from "how smart the model is" to "how much real value it can create."

However, under the Agent boom, a deeper technological question is emerging: the capability ceiling of Agents is essentially determined by pre-trained large models.

Pre-training: An Unavoidable "Ceiling"

If pre-training determines the model's ceiling, fine-tuning decides whether it can reach the ceiling. Professor Tang Jie from Tsinghua University clearly stated that pre-training allows large models to already possess world common sense knowledge and basic reasoning abilities—more data, larger parameters, and more saturated computation remain the most efficient methods for scaling base models.

This judgment is not an isolated case. The industry is forming a consensus: pre-trained large models are the ceiling of large model reasoning capabilities. Regardless of how reinforcement learning is refined in the post-training phase or how Agent frameworks are arranged, the model's reasoning depth, knowledge breadth, and generalization ability are fundamentally limited by the world knowledge and basic cognitive abilities injected during the pre-training stage.

The technical report of Ant Group's Ling & Ring 2.6 precisely summarizes this: "Large models are transforming from chatbots into agents. This transformation sounds like just a change in usage, but it actually completely rewrites the optimization goal of the model—A practical Agent must reason well." The foundation of reasoning ability lies precisely in pre-training.

Li Qiang, vice president of Tencent Cloud, has a clever analogy: "The large model is the engine, and the engineering chain is the process of turning the engine into a car. The engine determines the upper limit, and the engineering determines whether it can run, how far, and how smoothly." Agent frameworks, tool calls, and memory management—all these engineering capabilities are built upon the "engine." If the engine isn't good, even the most refined car won't go far.

The Prosperity of Agents Cannot Hide the Dependence on the Base Model

Current Agent capability evolution is verifying a core proposition: how complex an Agent can do depends on how deeply the base model can understand the world.

The open-source Agents-A1 model from Shanghai AI Lab was trained using high-quality long-range trajectory data across multiple domains and tasks, enhancing the model's understanding, reasoning, and instruction-following capabilities under long context conditions. The research team did not choose the shortcut path of "small model + complex arrangement," but returned to pre-training itself—improving the model's Agent capabilities from the foundation by using better data and more optimal training strategies.

These practices point to the same direction: the competition of Agent capabilities is ultimately a competition of base model capabilities.

The Limitations of Arranged Agents and the Rise of Native Model Agents

There are currently two technical routes for Agent development: one is arranged Agents, which are general large models plus scheduling frameworks like LangGraph and Dify, achieving task decomposition and tool calls through engineering means; the other is native model Agents, which directly embed planning and reflection logic during the training phase, offering stronger continuity.

Arranged Agents are flexible and easy to deploy, but face a fundamental dilemma: long tasks often lead to logical drift. When the task chain stretches to dozens or even hundreds of steps, the insufficient reasoning ability of the base model causes the Agent to "deviate" as it progresses. This is not a problem that the arrangement framework can solve—the framework can organize processes, but cannot improve the model's judgment ability.

A deeper change is taking place. With the rapid iteration of models, the Agents built by enterprises with substantial engineering resources may be covered by the new native capabilities of the next model version. This means that the barriers created by purely engineering arrangements are being constantly eroded by the continuous evolution of model capabilities.

The industry is moving from the era of training models to the era of training intelligent agents, but "model architecture and training data are still important"—environment design, infrastructure, and evaluators have entered the core circle, but they have not replaced the core position of the model itself.

Returning to the Model Core: The Final Form of Agents

The future form of Agent applications will ultimately return to a model-centric structure. This does not negate the value of Agent engineering, but reflects a clear understanding of the technical essence.

The engine determines the upper limit, regardless of how the Agent framework evolves, how rich the tool calls are, or how complex the memory system is, the Agent's reasoning ability, planning ability, and world understanding ability are always limited by its underlying base model. The scale of data, computing resources, and architectural innovations invested in pre-trained large models directly determine the intelligent height that the Agent can reach.

This means that for creators and users, the most worthwhile investment direction is not chasing short-term hotspots in Agent frameworks, but deeply understanding and continuously investing in base model capability building. Whether it is native Agent capability training for the model or continuous pre-training for Agent scenarios, it is more valuable in the long term than superficial engineering packaging.

In 2026, the AI industry is moving from "acquiring intelligence" to "organizational intelligence." But no matter how the organizational form changes, the model remains the source of intelligence. Agents bring the model into the environment and create productivity, but the model is the one that "can think." Without Agent capabilities, large models will remain in the theoretical learning stage—conversely, without a strong pre-trained base, Agents will stop at shallow execution and fail to reach true intelligence.

03     From Copilot to Agentic: New Programming Paradigms Under AI-Native Transformation

Software development is experiencing a paradigm shift as significant as the invention of the graphical interface. According to Gartner's definition, an enterprise-level AI coding agent is an autonomous or semi-autonomous software engineering solution that can perceive context, convert human intent into multi-step plans, and execute and validate these steps in code, tests, and related engineering products. From "AI-assisted programming" to "Agentic programming"—AI is no longer a passive code completion plugin but an independent developer with environmental awareness, autonomous planning, tool invocation, and self-correction capabilities.

Enterprise Value: From Individual Efficiency to Organizational-Level Productivity

For enterprises, Agentic programming is not just about making engineers write code faster. Its true value lies in reshaping the entire software production system.

Gartner predicts that by 2028, asynchronous AI coding agent workflows will increase software engineering team productivity by 30% to 50%, far exceeding the 0% to 20% improvement brought by AI code assistants in 2025. As of April 2026, the annualized scale of the global enterprise-level AI code intelligence market has reached between $9.8 billion and $11 billion. Anthropic's report further confirms this trend: 80% of leaders say that intelligent agent investments have already brought measurable financial impact, and 88% expect more returns in the future.

The real enterprise value comes from organizational-level Agentic deployment, not individual-level. At the end of 2025 to the beginning of 2026, Stripe, Ramp, and Coinbase almost simultaneously publicly disclosed their internal Coding Agents—Minions, Inspect, and Cloudbot. These three companies developed independently but eventually converged to almost the same architecture: the Agent is no longer "used by one person in the terminal," but "triggered by the entire team via Slack or GitHub Issue at any time." You need sandbox isolation execution environments, you need the Agent to resume previous work after interruptions, and you need to prevent one user's uncontrolled loops from burning up the company's model budget.

As IDC pointed out: The real difference is not whether AI coding tools are introduced, but whether the enterprise has a comprehensive plan for platform engineering, governance capabilities, and developer role transformation. Organizations that only pilot intelligent agents in local scenarios will find it hard to unlock scalable value; and those that build Agentic AI as an enterprise-level capability are more likely to gain long-term advantages in speed, quality, and innovation.