In late August, DeepSeek officially released its new model V3.1, which the company defines as the 'first step toward the age of AI agents.' Following the announcement, application-end sectors such as AI agents, information technology innovation, and gaming collectively strengthened in the market, with multiple companies hitting their daily limits. The market reacted enthusiastically to this narrative.

From last year's battle of large models to today's agent-centric narrative, the AI industry's focus is undergoing a clear shift: Whether models will become even smarter is no longer the only suspense (suspense); greater attention is now being paid to whether AI can truly complete tasks on behalf of humans.

01. Starting with a Release

The most notable change in DeepSeek V3.1 is that it integrates 'thinking' and 'execution' into a single model. This 671B-parameter model supports both thinking and non-thinking modes, which users can switch between using a 'Deep Thinking' button. The context window has been expanded to 128K, and the API now supports function calls in strict mode while remaining compatible with Anthropic's interface format, enabling integration with agent frameworks such as Claude Code. Official data shows that V3.1 demonstrates significant improvements over its predecessor in code repair benchmarks (SWE) and complex task testing in terminal environments, while also outperforming previous reasoning models in multi-step reasoning and complex search tasks. In terms of thinking efficiency, V3.1 reduces output tokens by 20% to 50% while maintaining task performance comparable to previous reasoning models.

Beyond technical parameters, two details are particularly noteworthy. First, DeepSeek added 840B tokens to V3.1's external training, with exponentially increased investment in long-context expansion, clearly betting on agent workloads. Second, according to third-party community evaluations, V3.1's programming benchmark scores have reached the highest level among non-reasoning models, with unit costs significantly lower than those of overseas counterparts. The combination of cost efficiency and tool-calling capabilities is precisely what agents need for large-scale deployment.

02. From 'Answering Questions' to 'Getting Things Done': What Agents Change

To understand the logic behind this market trend, we must first distinguish between two concepts: chatbots and agents.

The AI products familiar to us over the past two years are essentially 'question-and-answer machines': users ask questions, and the model responds. Conversations unfold through multiple rounds of Q&A, but each response is independent, and the model does not truly interact with the external world. Agents are different. Their core capability is 'execution': after receiving a task, they can independently plan steps, call tools, access databases, operate software, and verify results until the entire task is completed. If a Q&A model 'talks the talk,' an agent 'walks the walk.' A common analogy is booking a flight: a Q&A model can only provide flight information, while an agent will query, compare prices, place the order, and correct errors autonomously.

This capability leap is supported by the maturation of several infrastructure components. At the interface level, the MCP protocol unifies communication between models and external tools, allowing agents to use various services as easily as calling local functions. At the orchestration level, multi-agent collaboration frameworks enable complex tasks to be divided among specialized agents for parallel processing. At the evaluation level, observability and assessment systems are gradually being established, transforming agents from demos into monitorable and traceable production systems. The availability of these foundational capabilities is the technological prerequisite for 2026 being dubbed the 'Year of AI Agents' by multiple institutions.

03. Three Signals: Why the Inflection Point Is Now

Zooming out, the agent industry's boom is traceable, with three key signals converging in 2026.

The first signal is the steep adoption curve. Gartner predicts that by the end of 2026, 40% of enterprise applications will have built-in task-oriented AI agents, up from less than 5% in 2025—an nearly eightfold increase within a year, which Gartner considers one of the most aggressive enterprise technology adoption forecasts in history. Surveys show that 17% of organizations have already deployed AI agents, while 42% plan to do so within the next 12 months. IDC forecasts that by 2029, more than 1 billion AI agents will be deployed globally, for comparison, currently, there are only a few million mobile apps across all global app stores combined.

The second signal is the emergence of viable business models. Agents are no longer stuck in the proof-of-concept stage; use cases in customer service, office automation, e-commerce, and marketing have developed replicable deployment patterns, with monetization shifting from pure subscription models to hybrid 'subscription + usage-based + revenue-sharing' pricing. Token consumption is becoming a core metric for measuring enterprise AI maturity. Goldman Sachs estimates that global token consumption could be 24 times higher in 2030 than in 2026, with enterprise agents potentially driving a 55-fold increase in token demand by 2040.

The third signal is policy and standardization support. The 2026 government work report included 'agents' for the first time, setting a goal of over 70% agent application adoption by 2027. The 'Implementation Opinions on the Standardized Application and Innovative Development of Agents' identified 19 encouraged development directions. In late June, seven national standards under the 'Artificial Intelligence—Agent Interconnection' series were released, covering key areas such as overall architecture, identity management, interaction, and tool-calling, clearing standardization hurdles for cross-system collaboration.

04. Industrial Chain Relay: Models, Platforms, and Applications

The agent industrial chain can be roughly divided into three layers, each with its own business logic.

The model layer is the foundational infrastructure. Iterations of domestic models such as DeepSeek V3.1, GLM-5.2, and Kimi K2.7 have shifted from general-purpose Q&A to developer tools and enterprise workflows, with enhanced Agentic Coding capabilities being particularly critical. Guangfa Securities argues that the core driver of this AI industrial chain cycle is the leap in model programming capabilities, with programming tools evolving from Assist in completion (assistive completion) to autonomous collaboration. Competition at the model layer hinges on a combination of performance, cost, and openness, with open-source models driving down development thresholds through cost advantages.

The middle layer consists of agent development platforms and orchestration frameworks. In the first half of the year, at least 2,770 application-oriented large model bidding projects were awarded nationwide, with nearly 30% related to agents and around 180 involving agent development platforms. These platforms address the question of 'how to build agents': enterprises can deploy business-specific agents without training models from scratch by configuring workflows, integrating tools, and defining permissions on the platform. This is a 'selling shovels' business and currently the most active area for deployment.

The application layer offers the greatest profit potential. Office collaboration, customer service, content marketing, and software development are the fastest-growing agent penetration scenarios. In the office collaboration space, commercialization of AI assistants in mainstream productivity software is accelerating. In the enterprise market, service models delivering results based on 'token costs + management fees' are gaining traction. One company has scaled agent operations across four scenarios—performance advertising, brand advertising, content e-commerce, and short-form drama production—generating over 100 million yuan in revenue in its first year. As model-calling costs continue to decline, profit margins in the application layer are expanding.

05. Acceleration in the Chinese Market

The Chinese agent market is characterized by rapid deployment, strong policy support, and diverse scenarios.

Data shows that China's enterprise-grade AI agent market was worth approximately 8.6 billion yuan in 2024, surging to 21.2 billion yuan in 2025, and is expected to reach 44.9 billion yuan in 2026 and 332 billion yuan by 2029, with a compound annual growth rate exceeding 100%. According to IDC, the number of active enterprise-grade AI agents in China will grow from nearly 2 million in 2025 to about 5 million in 2026.

Additionally, nearly 30% of bidding projects nationwide in the first half of the year were agent-related, indicating that demand has shifted from 'trying out text generation' to 'embedding agents into real workflows.' At the 2026 China Internet Conference, Chinese Academy of Engineering academician Wu Hequan predicted that by 2030, the proportion of dialogue-based token traffic in China will drop from about half in 2025 to 12%, while traffic for agents and their services will rise to 75%, marking a complete reversal of AI's 'talking' versus 'doing' roles.

06. Opportunities and Risks Coexist

The opportunities brought by agents are structural. For enterprises, agents are the key to translating two years of AI pilots into productivity gains. For investors, the exponential growth in token consumption means the entire industrial chain's pie is expanding, with incremental gains at every layer—from models to platforms to applications.

However, risks must also be acknowledged. First, hallucination and reliability issues persist; strong reasoning modes may generate confident but incorrect multi-step outputs, requiring human review in critical decision-making scenarios such as finance and healthcare. Second, most agent applications are still transitioning from proof-of-concept to production deployment, with the breadth and depth of commercial viability yet to be proven. Third, model pricing adjustments, such as DeepSeek's announcement in September to adjust service prices and eliminate nighttime discounts, hint at cost pressures, suggesting that the model layer's price war may not be sustainable. Fourth, standards and governance are still evolving rapidly, and true cross-system agent interoperability will take time to materialize.

The age of agents has begun. The shift from 'conversation' to 'execution' may seem like a mere two-word change, but it represents AI's transformation from content generator to production tool. Whether this transformation delivers on its promise depends on the synergy of model capabilities, infrastructure, and application scenarios—as well as how long the market is willing to wait to validate this narrative.

- End -