Last year, OpenAI founder Sam Altman predicted at the first OpenAI Developer Conference that in the future, everyone in various industries could have an AI Agent. Bill Gates also wrote a lengthy blog post about AI Agents, stating that they will fundamentally change the way humans interact with machines and disrupt the entire software industry. Following the intensive emergence of large models, the trend of AI Agents has swept in.
Domestic tech giants have also launched their own AI Agent platforms, such as DingTalk’s AI PaaS, Baidu’s Smart Cloud Qianfan Large Model Platform, ByteDance’s Kouzi Space, and iFlytek’s Xingchen Intelligent Agent Platform, among others. The future is here, and with the continuous “evolution” of AI large models, the era of the explosive growth of AI Agents has already begun.
01 AI Agents: The Cambrian Explosion
1. What is an AI Agent?
As a new phenomenon, AI Agents are still in the exploratory stage from theory to application. In June 2023, OpenAI’s application research director Lilian Weng proposed:
Agent = LLM + Memory + Planning Skills + Tool Usage
In 2024, at the AI summit hosted by Sequoia Capital, Andrew Ng stated that an Agent should possess four main capabilities: Reflection, Tool Use, Planning, and Multi-Agent Collaboration.
It is evident that industry experts have a similar understanding of AI; an Agent can be simply understood as an intelligent entity capable of perceiving its environment, making autonomous decisions, and executing complex tasks. At this stage, Agents are primarily driven by large language models (LLMs), with memory, planning, and tool capabilities as key modules.
2. The Explosion of AI Agents
During the Cambrian period, approximately 540 million years ago, most animal phyla began to appear. This period lasted for about 20 to 30 million years and saw the emergence of many higher organisms, forming the basic patterns of species diversity, marking a period of explosive life on Earth. Similar to the Cambrian explosion of life, the emergence of experimental products like BabyGPT, AutoGPT, and Generative Agents has led to a similar explosion of Agents in the latter half of the large model era. The difference is that Cambrian life was carbon-based, while current AI Agents are silicon-based.
Like the evolution of life, the future world of Agents will see an increasing number of applications at the application layer. They will continuously upgrade and iterate, “evolving” into more complex intelligent forms. Although currently, the capabilities of Agents are still somewhat limited, tools like HuggingGPT have already demonstrated their ability to use tools in specific modules. As technology matures and progresses, Agents will inevitably evolve like humans, ultimately being able to think, execute, and autonomously solve various problems.
02 China vs. USA: Current Development of AI Agents
1. China: Giants Lead, Startups Surge
There is a saying in the investment circle: “Out of ten AI applications, five are office Agents, and three are AIGC.” As the recognized best carrier for LLM implementation, the development momentum of Agents in China can be described as a hundred boats vying for the current, bustling with activity.
First, the giants lead the way, taking the initiative. The aforementioned DingTalk, Baidu, and ByteDance primarily operate on a platform model, providing infrastructure for small and medium-sized companies, similar to OpenAI’s GPTs.
The AI PaaS includes three parts: model training platform, model scheduling platform, and plugin development platform. The bottom layer provides AI infrastructure support, the first layer opens various general large models and plugins, allowing enterprises to call large models for parameter tuning to create their own exclusive models; the second layer is the model scheduling platform, which includes context memory, intelligent planning, template management, general capabilities, and prompt tools. Using the tools and capabilities provided by these two layers, enterprises in the ecosystem can develop a variety of AI applications. Additionally, DingTalk also offers different scenario applications and industry solutions at the third layer.
Baidu’s Smart Cloud Qianfan Large Model Platform is also composed of AI infrastructure, basic management platform, general large model training, prompt engineering, model fine-tuning, and other functions. It has its own strengths compared to AI PaaS. In comparison, Baidu’s large model training is more detailed and rich in functionality, while DingTalk focuses more on simplifying processes to lower the user threshold.
In addition to the large companies, startups focusing on the application layer are also keenly eyeing Agents. They often have the agility to pivot quickly, especially those focusing on vertical fields, which have a greater opportunity for rapid innovation and launching corresponding products and solutions.
2. USA: Clear First-Mover Advantage, Scene Explosion
In the entire AI field, the USA has a clear first-mover advantage, and the same goes for Agents. They are not only ahead in technology but also have a significantly richer array of application scenarios. For example, Microsoft’s AutoGen, OpenAI’s GPTs, programming Agent Devin, customer service Agent4, and retail Regie.AI.
AutoGen allows multiple LLM agents to solve tasks through conversation. LLM agents can play various roles, such as programmers, designers, or combinations of various roles, solving tasks through dialogue.
OpenAI’s GPTs allow anyone to create a customized version of ChatGPT, which can help you learn the rules of any board game, assist your child with math, or design stickers. Anyone can easily build their own GPT without coding knowledge.
Devin can plan and execute complex engineering tasks involving thousands of decisions, recalling relevant context at each step, and learning and correcting errors over time. It can not only automate tasks but can even write an entire application by itself in a matter of minutes.
Agent4 can understand natural language, engage in smooth conversations with users, and provide personalized services based on user needs and preferences. It can handle calls for various brand skills, accurately answer customer inquiries based on a knowledge base, and quickly grasp new company policies and business operations.
Regie.AI can create custom sales sequences, write highly personalized emails, and store relevant sales content. It also integrates with leading sales engagement platforms (such as Outreach.io, SalesLoft, and Hubspot), reducing the time required to bring messages to market and achieve initial results. Regie.AI is well-suited for retail and e-commerce, shaping a customized and data-rich shopping experience aligned with core retail goals and objectives.
03 How Far Are AI Agents From Us?
1. The “Brain” Is Not Yet Perfect
In the AI field, large models are seen as the brain of Agents, and the combination of “multi-modal large models + Agents” is considered a feasible path to AGI. Agents can adapt to changing application environments through continuous learning, capable of handling known multi-modal tasks and quickly responding to unknown multi-modal tasks. However, people also have higher expectations for Agents, hoping they can possess genuine creativity by autonomously exploring the environment to discover new strategies and solutions.
But the reality is that the current large model “brain” is not very complete and is insufficient to support a greater degree of Agent functionality. This is also why large models have not yet achieved AGI (Artificial General Intelligence); the development of multi-modal capabilities (such as image and video recognition and generation) is still ongoing (e.g., models like Sora are not yet perfect) and represents a significant bug.
2. The “Limbs” Are Not Strong Enough
Whether it is the support of MCP or A2A protocols, or the interfaces of existing websites and apps, they are akin to the limbs of Agents. A healthy tool ecosystem is crucial for the development of AI Agents. Over the past year, the development of the Agent tool ecosystem has been rapid, including browsers, programming IDEs, vector databases, etc., attracting a large number of entrepreneurs. The tools that Agents can call upon are increasing. However, it will still take a considerable amount of time to transition from quantitative to qualitative changes.
3. Lack of a Universal External Framework for Agents
A unified underlying development framework can provide shared services and functions for AI Agents, including various tools and libraries for data processing, model training and testing, and tools for deploying and monitoring AI Agents. It can help developers quickly develop and deploy AI Agents without having to build all the infrastructure and functionalities from scratch.
Although there are already some open-source frameworks available, there is still a long way to go from being usable to being user-friendly.
4. A Flourishing Ecosystem Is the Key
In the field of AI Agent development, there are currently several large platforms and companies both domestically and internationally that are conducting effective research and continuously launching applications, such as DingTalk, Baidu, ByteDance in China, and Microsoft, OpenAI abroad. However, compared to a situation where “one (or a few) stands out,” forming a healthy ecosystem of “a hundred flowers blooming” for AI Agents is undoubtedly more important. This objectively requires more small and medium-sized software service providers to participate and contribute.
Due to their deep understanding of specific business scenarios, small and medium-sized software service providers can often develop AI Agents that better meet user needs, thereby improving the quality of AI Agents. For example, a software service provider focused on the e-commerce sector may develop an AI Agent that provides personalized recommendations based on users’ shopping history and preferences. Additionally, because small and medium-sized software service providers are relatively smaller in scale, they can often experiment with new technologies and methods more quickly, thus driving innovation in AI Agent technology.
04 AI Agents Are Revolutionizing toB Software
1. The Revolution of the toB Software Ecosystem Has Arrived
AI Agents are gradually changing the toB software industry. In the future, enterprises will increasingly rely on integrated platforms such as DingTalk, Feishu, and WeChat Work, which provide one-stop solutions, including clocking in, attendance, business process approvals, OA, ERP, CRM, and more. These functions are all “built into” the platform, meaning they are integrated within the same application or platform, allowing users to complete all their work without leaving the platform. This design makes the user experience smoother, reduces costs, and increases efficiency. This advantage will inevitably lead to a decrease in independent entry points for toB software. After all, no one wants to switch back and forth between multiple software applications, which is inefficient, time-consuming, and costly.
2. AI Agents Accelerate the Elimination of toB Software
The toB software that is being accelerated towards elimination includes: simple data analysis, standard process types (OA, business, marketing, finance, etc.), execution automation types (RPA software), light consulting, and traditional education.
1) Data Analysis Software
This type of software can process, analyze, and visualize massive amounts of internal and external data for enterprises, helping them make more informed decisions. However, with the continuous development of large model technology, some AI platforms can begin to provide more intelligent data analysis and prediction services, rendering this type of toB software obsolete and facing replacement.
2) Standard Process Software
The powerful data processing and learning capabilities of large models can automate the handling of repetitive, clearly defined tasks by learning and analyzing a large amount of office data, thus saving time and effort spent on simple repetitive manual processing. Therefore, like all highly standardized professions, standard process software such as OA (Office Automation), business approval flows, and marketing automation also face the risk of being replaced.
3) Execution Automation Tools
This type of tool (RPA) primarily excels at simulating human operations on computers, thereby replacing human labor in various repetitive, clearly defined tasks. For example, in finance, RPA can automatically execute accounting processes, data entry, invoice verification, etc.; in customer service, RPA can automatically reply to customer emails and handle complaints using natural language processing technology. However, its drawbacks are also evident, such as its inability to handle complex and highly variable processes, and tasks requiring human decision-making and judgment are not well-suited for RPA.
4) Light Consulting Service Software
This type of software can help enterprises understand industry dynamics, market trends, and competitive situations, providing decision support. However, with the development of AI technology, large models provide intelligent industry analysis and consulting services that are actually superior and more convenient. Therefore, in the future, not only software but even the entire light consulting industry will be significantly impacted by AI large models.
5) Traditional Education Software
Compared to traditional online education software, such as programming education and English tutoring, AI has more obvious advantages. For example, AI large models can automatically generate teaching plans and provide more personalized and precise teaching content, thereby reducing reliance on human teaching.
3. Three Types of Software Will Improve
These three types of software are: complex business management, software with scarce data, and industry management software.
1) Complex Business Management Software
Complex business management software such as ERP, WMS, and TMS plays an important role in enterprises, helping them optimize resource allocation, manage inventory, and monitor transportation processes. However, complex business management software often needs to handle large amounts of data and complex processes, requiring customized development based on the actual situation of the enterprise. AI technology can provide auxiliary functions to some extent, but complete replacement is still challenging.
2) Software with Scarce Data
Some industry-specific software, such as precision manufacturing, discrete manufacturing, and new drug development management software, typically need to process large amounts of data and complex algorithms. The data and algorithms in these software are often scarce resources, owned by only a few enterprises. Additionally, in handling highly specialized data and knowledge, human expertise and experience still play a crucial role, making the threshold relatively high and difficult for AI large models to completely replace.
3) Industry Management Software
Industry management software, such as equipment production management software and new drug development management software, plays an important role in certain specific industries. This type of software can help enterprises standardize, streamline, and regulate management in research and development, production, quality, and materials. Due to the significant differences in business and management models across different industries, and the complexity of industry rules and knowledge that these software typically need to handle, it is challenging for AI technology to develop universal industry management software for each sector.
05 toB Software: Stand Out or Be Eliminated
1. Stand Out or Be Eliminated
AI large models are gradually devouring the market share of toB software, which is an increasingly evident reality. However, this does not mean that toB software can only sit idly by. Due to their deep understanding of the industry, toB software companies can customize AI Agents based on the characteristics and needs of the industry. These AI Agents can better meet the specific needs of the industry (such as healthcare, finance, etc.), providing more precise and effective services. For example, a software company serving the healthcare industry may develop an AI Agent that helps doctors analyze medical records and provide diagnostic suggestions, which is something that other “outsiders” find difficult to achieve.
In contrast, tech giants like Google and Amazon, while having strong capabilities in AI technology, cannot or are unwilling to delve into every specific industry to understand its characteristics and meet its specific needs due to considerations of time and resource investment. This is a key pathway for the toB software industry to thrive amidst challenges.
2. If You Can’t Beat Them, Join Them
As AI has developed to this point, forward-thinking toB software practitioners must establish a mindset of collaboration and partnership. They need to have the insight to “join forces” when they cannot compete, fully leveraging available resources, whether it is DingTalk’s AI PaaS or Baidu’s Smart Cloud Qianfan Large Model Platform, to develop and deploy AI Agents quickly without worrying about underlying infrastructure and operational issues. Moreover, these toB software can gain more customers through services from these platforms, increasing revenue while collecting more data, which is beneficial in every way. In a challenging world, survival is not easy, and taking advantage of such opportunities is not shameful.
Steve Jobs once said, “People don’t know what they want until you show it to them.” AI Agents are demonstrating their unmatched capabilities and efficiency to an increasing number of people and potential users across various industries. It is time for the toB software industry and its professionals to prepare to take action, whether to keep pace with the times and seize the opportunities of the AI era or to adjust their direction and make alternative choices. In any case, it is time for everyone to get moving and do something. What do you think?