How to Blend Process‑Oriented and Agent‑Centric AI into a Hybrid Intelligent Pipeline

Two Paradigms for Building LLM‑Based Agents

In the LLM era, two technical streams dominate:

Process‑oriented intelligence (e.g., Coze) treats the workflow as a deterministic graph of nodes; each node is a sub‑task that the LLM fulfills. Advantages: explicit dependencies, predictable state transitions, interrupt‑and‑resume, full execution trace.

Agent‑centric intelligence (e.g., DeerFlow) equips a single agent with perception, reasoning and planning abilities, allowing it to adapt its plan dynamically, incorporate feedback, and explore multiple solution paths.

Why a Hybrid Architecture?

Pure pipelines guarantee high success rates but suppress creativity; fully autonomous agents are flexible but can be unstable. A hybrid design combines deterministic execution with autonomous planning, tool use, and memory.

Core Fusion Modules

Plan & Reflection

The planner generates a directed‑acyclic‑graph (DAG) of sub‑tasks, tool calls and dependencies from the user query. The executor runs the DAG. A reflection module monitors low‑confidence or blocked states and rewinds to the planning node to revise the graph.

Tool Use

Tools are invoked via carefully crafted prompts. The agent learns selection, parameterisation and composition of tools through reinforcement‑learning‑style feedback, improving its tool‑selection policy over time.

Memory Mechanisms

Memory bridges instant reactions and long‑term evolution. It is organised into:

Retrieval memory (RAG) : external knowledge retrieval to reduce factual conflicts.

General memory : pre‑/post‑training storage of universal knowledge.

Rule memory : prompt or RL‑driven constraints (e.g., chain‑of‑thought, JSON output).

Short‑term memory : session‑level cache (Redis + summarisation + vector store) for immediate responsiveness.

Long‑term memory : accumulated user profiles for personalised behaviour.

Implementation Challenges

Data : acquiring high‑quality, diverse SFT and RL data.

Training : designing reward functions that balance end‑to‑end performance with stable convergence.

Engineering : distilling models, assembling multi‑size ensembles, and decomposing tasks for production deployment.

Key References

Baidu AI Search Paradigm (arXiv): https://arxiv.org/pdf/2506.17188

Anthropic Multi‑Agent Research System: https://www.anthropic.com/engineering/multi-agent-research-system

Illustration

Code example

本文
约5300字
，建议阅读
5
分钟
“RAG是面向LLM的搜广推，Memory是面向Agent的数据飞轮”