AI Agent: The “disruption” and “latecomer advantage” brought by big models

I have been thinking about one thing recently: will the development of large models lead to fundamental changes in existing AI Agent products, or even replacements?

For example, with the development of multimodal capabilities of large models, when a single large model can integrate multimodal interactions and handle text, images and voice at the same time, the multi-agent systems designed specifically for this purpose in the past will become obsolete.

This kind of disruption can easily lead to "latecomers catching up". Faced with such pressure, what kind of mentality should practitioners in the field of AI Agents adopt and how should they prepare?

Next, let’s talk about this topic.

Valuable and promising AI Agent products must be the product of the deep integration of big model capabilities and engineering accumulation.

The essence of the development of large models is to continuously lower the threshold of general intelligence. Therefore, the intelligence threshold of AI Agents will also be gradually lowered.

In this context, in addition to the impact of the multimodal capabilities mentioned at the beginning, the upgrade of the general task processing capabilities of large models has gradually enabled it to have the ability to plan complex tasks, which has impacted the original "task decomposition + tool call" implementation architecture of the Agent.

For example, in scenarios such as text summarization and basic data analysis, if large models can be directly called to meet the needs, the tool chain integration of traditional agents will no longer be necessary.

Therefore, agents that rely solely on “basic functions” do run the risk of being directly integrated into large models.

In addition, as the real-time and reliability of large models improve, the engineering optimization value of agents in the past has also been greatly reduced. For example, in the industrial application field, by deploying lightweight models on edge nodes, millisecond-level responses have been achieved.

In fact, the big model does not subvert the Agent itself, but the "inefficient old paradigm". Its essence is the balance between "universal intelligent base and vertical scenario requirements".

When it comes to coping strategies, we must first get rid of the misunderstanding of "head-on competition with big models" and then look at it from different time perspectives:

• Short-term: Use agile architecture and lightweight experiments to quickly verify scenarios and reduce technology iteration costs;

• Mid-term: Deepen industry know-how and become a "vertical solution provider for large model ecosystem";

• Long term: I believe that the intelligent industry will continue to differentiate and focus on "human needs that are difficult for machines to replace."

We can discuss this in more detail by delving deeper into the technical architecture and scenarios.

1. Technical architecture: building a flexible and scalable "Lego-style" system

First, we can shift the Agent technology architecture to "large model base + scene adaptation plug-in", using the design concept of large model as the "base" and scene solution as the "plug-in".

That is, the big model is responsible for general intelligent tasks such as natural language understanding, complex reasoning, and knowledge generation, while the scenario plug-in handles engineering logic such as real-time data access, tool calling, process orchestration, and domain constraints.

In this way, when the large model version is upgraded, it is only necessary to update the base interface definition, and then the plug-in will adapt to the new output format, avoiding the problem of "a single move affecting the entire system".

Therefore, the interaction protocols and interface specifications between the large model base and the plug-in must be unified.

For most companies, there is no need to develop large models on their own. They only need to focus on scenario plug-in innovation and allow the Agent architecture to evolve towards "decentralization".

2. Deepening the scene: Building data and knowledge barriers

In the future, Agent competition will present a pattern of "base convergence and plug-in stratification".

Leading companies build "deep-water" barriers through in-depth industry knowledge, while small and medium-sized enterprises rely on general plug-ins to quickly implement "shallow-water" applications, ultimately forming an industrial ecosystem of "large models for inclusive intelligence, and scenario knowledge to define value."

Therefore, under the architecture of "large model base + scene adaptation plug-in", the depth of knowledge in vertical fields will become the key to building differentiated barriers for Agent products.

Specifically, it can be viewed from the following three aspects:

• Integration and construction of industry data: Make good use of proprietary data as assets. For example, industries such as medical and financial industries need to accumulate proprietary data such as case records and transaction records, and actively introduce external high-quality data assets to integrate data and form unique data asset advantages.

• Accumulation of vertical industry know-how: Convert industry knowledge, compliance requirements, operating procedures, etc. into knowledge graphs. For example, the accumulation of knowledge of extreme scenarios such as industrial equipment control rules in weak network environments requires long-term practical accumulation, which cannot be replicated by competitors in the short term.

• Safe decision-making design for human-machine collaboration: Identify key decision nodes in each scenario, such as the partner second-instance system in the "legal document generation scenario" and other organizational experiences, fully accumulate such node information involving implicit decision-making, and form a safe decision-making advantage.

In the wave of technology, the real risk is never "being disrupted" but "refusing to change."

Therefore, AI Agent practitioners need to view the iteration of large models as "capability upgrades" rather than "starting from scratch", using large models to raise the "cognitive ceiling" of agents, while using the ability to explore scenarios to build a "landing moat".

Ultimately, only Agent products that can find a balance between "universal smart base" and "vertical scene depth" will take the lead in disruption.