In 2025, How Will AI Agents Transform?
Moving from traditional AI agents to personal foundation agents.
2024 is regarded as the year of AI applications by the industry. In 2025, agents are widely expected. Last week, Google officially released its latest large model, Gemini 2.0 series, claiming that this is their most powerful artificial intelligence model to date, “designed specifically for the agent era.” How to use “intelligent agents” as the core driver to break the limitations of traditional human-computer interaction has become a core issue of heated discussion in the industry.
At the Geek Park IF2025 Innovation Conference, Zeng Xiaodong, founder and CEO of Boundless Ark, in his keynote speech titled “Besides Finding an AI as a ‘Lover’, What Else Can AI Do?”, deeply discussed the future development direction of AI agents, especially how to promote AI from a single-task assistant to a personalized and emotional intelligent partner through foundation agents.
**I. How did AI agents develop?**
1. The development route of AI agents
– Agents first appeared in human vision nearly 20 years ago with AlphaGo. In reinforcement learning, agents interact with the environment in a large number of chess games and learn the task of playing Go. But these agents can only handle single tasks. So after AlphaGo, agents did not receive more attention for a long time until the emergence of large models.
– Taking language foundation models as an example, they can handle many tasks, including middle and long-tail tasks. Soon, many basic agent frameworks grew on LLM, and many Prompt Agents also appeared. That is, by writing prompts to give agents a certain role and configuring some callable tools. Currently, there are more than 700,000 Prompt Agent applications worldwide. Now, as long as you open any large model app, there will definitely be a tab about intelligent agents. These agents are collectively referred to as Prompt Agents or Baby Agents because they are essentially some general capabilities of large language models, just materializing their functions through writing prompts.
– In the future, AI agents will have two deep waters: expert agents and personal agents.
2. Expert agents
– When agents enter vertical fields, more professional agents are needed. Prompt Agents directly applying general models cannot meet the professionalism requirements of vertical fields. The task pass rate of general models in vertical fields is often less than 50%. So expert agents are needed to deeply couple the model with vertical field data and professional business processes to form highly professional AI agents.
– For highly complex tasks, a multi-agent team can be formed to solve particularly difficult propositions. For example, the product launched by Boundless Ark in the first half of the year is used to攻克难题 in the field of pharmaceutical research and development. There are 18 specialized agents in this product. The model behind each agent is different. The 18 agents can communicate with each other in natural language, write code, call medical tools and models, and automatically correct errors to handle highly complex problems.
– For the level of expert agents, the key to generating a business model is the professionalism of agents in that vertical field. Whether adopting a single-agent solution or a multi-agent solution, it must effectively play a role in reducing costs and improving efficiency.
3. Personal agents
– When agents enter the personal field, they can not only help users improve production efficiency but also provide more emotional value. Personal field agents not only exist in mobile phones or computers but will also be搭载 on more terminals, such as glasses, smart speakers, future humanoid robots, and more new types of smart hardware.
– There are still many core problems to be solved between foundation models and applications, such as interaction experience, personalized memory, execution ability, and so on. The team of Boundless Ark believes that what the personal field needs is definitely not a traditional agent but a foundation intelligent agent, that is, Personal Foundation Agent.
**II. Three elements of foundation intelligent agents: interaction, memory, and skills**
1. Interaction
– Personal AI applications need拟人化 and real-time voice and visual understanding interaction capabilities. However, the traditional method uses a “three-stage” serial link to achieve audio and video interaction, that is, first connecting an automatic speech recognition (ASR), then a large model LLM, and finally connecting a text-to-speech synthesis service (TTS). This method has three fatal problems: high latency, rigid interaction, and lack of emotion.
– End intelligence compresses and deploys the model to the end side to reduce interaction latency, but there will be problems of power consumption and a decline in model intelligence.
– The interaction ability of agents needs to be a completely open, very low-latency, visual understanding-capable, emotionally expressive, and able to drive software and hardware carriers.
– Boundless Ark has independently developed an
