Future Trends of AI Agents: Multi‑Agent Systems, Human‑AI Collaboration, and Multimodal Embodied Intelligence

The piece originates from Knowledge Map 3.0’s Agent module and presents insights from Qi Xiang, the NLP algorithm lead at Ant Financial and a PhD from the Chinese Academy of Sciences, who focuses on ToB serious‑domain agent research such as complex task reasoning, scenario evaluation, and system evolution.

Agent development is expected to advance along three primary axes: multi‑agent systems, human‑AI collaboration, and multimodal/embodied intelligence.

First trend – Multi‑agent: Two technical streams are identified. The workflow‑based stream, exemplified by BabyAGI and Task Weaver, relies on fixed pipelines that tightly orchestrate multiple agents, offering high stability and reliability for well‑defined tasks like plugin invocation and data analysis.

The conversation‑based stream, represented by MetaGPT and AutoGen, adopts a dialogic, role‑driven cooperation model where agents interact more freely, suitable for exploratory scenarios such as games or complex projects, and even allowing humans to act as an additional agent in the conversation.

The Data Intelligence Knowledge Map 3.0 comprises 26 modules grouped into four domains:

Data Production: data integration, data warehouses, data lakes, vector databases, graph databases, data weaving, unified stream‑batch processing, Data+AI.

Data Consumption: data governance, A/B testing, metric systems, DataOps.

Data Modeling: graph neural networks, knowledge graphs, large‑model data quality, LLM fine‑tuning, Retrieval‑Augmented Generation (RAG), agents, LLMOps, AI infrastructure.

Data Applications: intelligent risk control, recommendation, customer service, ChatBI, AI search.

Second trend – Human‑AI Collaboration: Humans increasingly serve as supervisors and assistants; agents must render their reasoning transparently and be receptive to human feedback and correction.

Third trend – Multimodal & Embodied Intelligence: While large language models excel at textual dialogue, agents can be extended to perceive and act through vision, audio, and physical interaction, opening applications in robotics, smart glasses, and other embodied systems.

The article concludes by inviting readers to download the Knowledge Map for free, emphasizing its value to professionals across the data and internet industries.