The theory of an AI bubble is heating up. Why has this year become the explosive year for agents?

In the landscape of the AI industry in 2025, a contradictory "dual - focus drama" is unfolding:

On one hand, there is a volatile correction in global technology stocks. The Hang Seng Tech Index has seen a maximum pullback of over 15% in two months. Leading stocks like NVIDIA have experienced a "high - open, low - close" roller - coaster ride in a single day, and the view of "the burst of the AI bubble" has been on the rise.

On the other hand, there is a counter - trend explosion in the AI Agent track. The market size of the AI Agent in China has soared from 4.75 billion yuan in 2024 to 7.84 billion yuan, a growth of over 60%. Chinese and foreign giants such as Baidu, Tencent, and Microsoft have all placed their bets, and the number of implementation cases in fields such as healthcare, industry, and finance has increased exponentially.

When skeptics classify AI Agents as "a new packaging of the bubble", the real transformation in the industrial sector has provided the answer: In 2025, it is not a year of AI bubble frenzy, but the commercial starting year when AI Agents move from concepts to value realization.

01 From "Conversation Tool" to "Execution Entity"

An AI Agent is considered a "software program capable of autonomously understanding, planning, and executing complex tasks". Driven by large - language models, it can autonomously call tools and systems and complete high - level goals without step - by - step human prompts.

This is fundamentally different from traditional AI assistants. When you ask a traditional AI assistant "how to analyze industry financial reports", it will provide methods. However, when you give the same instruction to an AI Agent, it can autonomously crawl data, conduct cross - verification, and generate structured research reports, only requiring final review throughout the process.

The birth of AI Agents is not accidental but an inevitable result of the iteration of AI technology.

The emergence of ChatGPT in 2022 marked the beginning of the era of large models. However, early large models mostly remained at the "conversation demonstration" stage and were criticized as "talk - only AI".

With technological breakthroughs such as Chain of Thought (CoT) training and Retrieval - Augmented Generation (RAG), large models have achieved a leap from "understanding language" to "logical reasoning". The maturity of function - calling capabilities has equipped them with "hands and feet", enabling them to operate software and call data. Multimodal fusion technology has further enabled AI Agents to "see" images and "hear" voices, forming a complete closed - loop of "perception - thinking - decision - execution".

In 2025, large models represented by GPT - 4 and Gemini 2.0 have completely transformed, driving AI Agents from the laboratory to the industrial field.

The core value brought about by this evolution is the transformation of intelligence from "cost" to "productivity".

Robin Li, the founder of Baidu, hit the nail on the head at the Baidu World Conference 2025: "When AI capabilities are internalized as native capabilities, intelligence is no longer a cost but a productivity."

Zhai Feng, the Chief Technology Officer of IBM Greater China, believes that AI Agents will play an "empowering" role and become collaborators in the intensive workflow led by humans. Low - value and repetitive tasks will be automated, and the potential of humans in high - level fields such as strategic planning and creative innovation will be further unleashed.

Zhang Shaofeng, the founder and CEO of BaiRong, has subverted the traditional perception that "AI is a tool" and proposed the core concept of "silicon - based labor force". He advocates regarding AI Agents as "silicon - based employees" of enterprises rather than "software purchased once".

According to Gartner's prediction, the proportion of autonomous AI integrated into enterprise software will soar from less than 1% in 2024 to 33% in 2028. Meanwhile, over 15% of daily work decisions will be autonomously made by AI Agents.

02 Giants' Bets and Track Explosion

The explosion of AI Agents in 2025 stems first from the strategic bets and resource investments of global technology giants. This investment is not a blind follow - up but is based on clear expectations of commercial returns, forming a positive cycle of "R & D - implementation - profit".

At the Baidu World Conference 2025 in November, Baidu unveiled a significant achievement - the world's first commercially available self - evolving super AI Agent, "Baidu Famo". Drawing on evolutionary algorithms, it can compress the hundreds of millions of years of evolutionary processes in the biological world into just a few days and find "globally optimal solutions never discovered by humans" in fields such as transportation, energy, and finance.

Tencent has built a moat for AI Agents based on the WeChat ecosystem. Liu Chiping, the CEO of Tencent, has clearly stated that WeChat will ultimately launch an AI Agent to help its 1.4 billion monthly active users complete various tasks within the ecosystem.

Overseas giants have also been active: Microsoft has integrated AI Agents into Dynamics 365, helping Lumen reduce its annual costs by $50 million. ChatGPT Agent by OpenAI has covered 500 million users within half a year of its launch. It can automatically browse web pages, operate documents, and complete multi - step tasks.

In vertical tracks, the commercial value of AI Agents has blossomed in all aspects and has become a "must - have" for enterprises to reduce costs and increase efficiency.

In the healthcare field, Ambience Healthcare, an American AI enterprise, has launched an AI - driven clinical assistance system composed of six functional - module AI Agents, which has re - engineered the medical documentation process. Pilot data shows that the time doctors spend on documentation per week has decreased from 20 hours to 8 hours, satisfaction has increased by 65%, the efficiency of medical record generation has increased by 8 times, the processes of referrals and discharge summaries have accelerated by 70%, and the daily patient reception capacity of hospitals has increased by 22%.

In the retail sector, Walmart has transformed from traditional retail to a "data - driven intelligent platform" through its "role - driven AI Agent system" in 2025. The Agent - based closed - loop from product selection, inventory, sales to after - sales has reduced the operating costs of individual stores by 22% and increased the customer repurchase rate by 15%.

In the financial field, BaiRong and a consumer finance institution have jointly developed a "silicon - based employee" focused on post - loan voice quality inspection, which has been implemented in the industry. As the AI technology provider, BaiRong has used its "BaiRong Baigong" AI Agent construction platform and large - model technology as the "brain" and "skeleton" of the "silicon - based employee", which will reduce the cost of manual quality inspection by 60% and increase the overall post - loan operation efficiency by over 40%.

The enterprise - level market has become the main battlefield for AI Agents. According to the prediction of Haibi Research Institute, the market size of AI Agent applications in China will reach 1.09 billion yuan in 2025 and will exceed 10 billion yuan in 2027. The financial, manufacturing, and software Internet industries will be the top three application fields.

03 The Ecosystem Resonance Behind the Explosion

The counter - trend explosion of AI Agents amidst the controversy over the AI bubble is essentially the resonance of three factors: technological maturity, upgraded demand, and a perfected ecosystem. Its explosion is inevitable.

The breakthrough in technological maturity is the core prerequisite.

In 2025, the inference cost of large models has decreased by 90% compared to 2023, and the inference speed has increased by 10 times, completely solving the cost bottleneck for the large - scale application of AI Agents.

Meanwhile, the technological architecture has become standardized. Platforms such as AWS Bedrock Agent Core and Baidu GenFlow provide modular components, eliminating the need for enterprises to build from scratch. A retail enterprise completed the development of a supply - chain AI Agent in just three days through the ChatFlow platform, compressing the response time from 72 hours to 8 hours.

The improvement of computing power infrastructure provides underlying support. The orders for NVIDIA's H200 chips increased by 300% in the first half of 2025. Runze Technology's net profit in the first three quarters reached 4.7 billion yuan, a year - on - year increase of 210%. Its intelligent computing center provides full - stack services for the development of AI Agents.

The rigid demand of enterprises is the direct driving force.

Currently, cost reduction and efficiency improvement have become the core demands of enterprise operations, while traditional AI tools are difficult to handle complex business scenarios. The "non - invasive" solution of AI Agents exactly meets the demand - it can automate processes by simulating human operations without reconstructing the existing systems of enterprises.

The global skills shortage has also prompted enterprises to introduce "digital employees". After an e - commerce enterprise applied an AI Agent, its price - adjustment response was 40 minutes faster than its competitors, and its sales volume increased by 22%. The support of policies and capital has formed an ecological synergy: China's "Artificial Intelligence +" strategy and special policies in Beijing and Shanghai encourage implementation through model vouchers and computing power subsidies. In 2025, the financing amount in the global AI Agent track exceeded 66.5 billion yuan, and 80% of it flowed to enterprises with clear scenarios. Capital has shifted from "blindly chasing concepts" to "precisely investing in value".

Looking to the future, AI Agents will move from "large - scale" to "refined" development, opening a new era of human - machine symbiosis.

Technologically, the combination of quantum computing and multi - agent reinforcement learning will increase the decision - making speed of urban - level traffic scheduling by 10⁶ times. 6G communication will support microsecond - level collaboration, promoting the "zero - latency" evolution of the industrial Internet.

In terms of applications, some institutions predict that 60% of enterprises will use AI Agents as the core operational support in 2026. Knowledge workers will be equipped with an average of 5.2 dedicated Agents per person, and work efficiency will increase by 300%. The multi - agent teaching system in the education field will increase the coverage rate of personalized learning from 35% to 90%.

In terms of governance, the China Academy of Information and Communications Technology has proposed a "technology - management - law" trinity framework, and the EU's "Artificial Intelligence Act" has clarified the certification standards for high - risk Agents, promoting the evolution of AI Agents from "usable" to "trustworthy".

Of course, challenges still exist: communication delays in multi - agent collaboration, insufficient interpretability of decisions, and the talent gap are all issues that need to be urgently addressed.

However, these issues are an inevitable stage of technological revolution rather than fatal flaws. Different from the concept stocks with "no profit, no revenue" during the dot - com bubble in 2000, AI Agents in 2025 have formed a closed - loop of "technology - scenario - value". Deloitte data shows that 73% of enterprises that have deployed AI Agents have achieved cost reduction, and 58% have achieved revenue growth.

The wave of AI Agents in 2025 is, as Robin Li said, a turning point when "intelligence changes from cost to productivity". The controversy over the bubble is exactly an inevitable path for the maturity of the industry. When the tide recedes, those concept - based projects that rely on subsidies and lack value will be eliminated, while AI Agents that truly solve industry pain points and create quantifiable value will become the infrastructure of the digital economy.

2025 is not the end of the AI bubble but the starting point for AI Agents to reshape the global productivity landscape. A new era of human - machine collaboration has quietly begun.

This article is from the WeChat official account "Unicorn Observation". Author: Unicorn Observation. Republished by 36Kr with permission.