2. Background and Challenges

2.1 AI Agent System Isolation

Currently, many AI agents are isolated within their respective platforms, lacking unified communication protocols and interoperability. This lack of interoperability limits the overall effectiveness of AI systems, especially in scenarios requiring cross-platform collaboration.

2.2 Limitations of Single LLM

Most AI agents rely on a single language model (such as GPT-4 or Llama2). This presents a concentration risk; performance can be compromised when facing complex multi-step tasks or multilingual scenarios.

The limitations of a single LLM model prevent systems from flexible switching or parallel usage of multiple models, thus restricting their application scope and performance.

2.3 Insufficient Retrieval-Augmented Generation (RAG) Accuracy

Most AI agents use retrieval-augmented generation to retrieve information from a specialised knowledge base, but it is challenging to achieve perfect knowledge base refinement and accuracy as the information and documents might be irrelevant, outdated, or low-quality.

2.4 Lack of Event Tracing and Observability

Existing AI systems often lack source management for AI agent internal states and historical interactions. When system failures or reasoning anomalies occur, it is difficult to locate problems and replay events, increasing maintenance complexity and risks.

2.5 High Deployment and Collaboration Costs

Traditional AI systems typically require complex installation, configuration, and maintenance processes. The lack of user-friendly workflow orchestration tools, especially in multi-agent collaboration scenarios, leads to high development and maintenance costs.