Unlocking LLM‑Based Agents: Architecture, Challenges, and Future Directions

Introduction

With the rapid maturation of large language models (LLMs), AI agents built on LLMs have become increasingly visible. This article reviews essential knowledge about LLM‑based agents and discusses important application directions in the era of large models.

Overall Architecture of LLM‑Based Agents

The agent framework consists of four primary modules:

Profile Module : Describes background information of the agent, including demographic, personality, and social data. Generation strategies include manual prompt design, large‑model‑generated profiles, and data‑alignment prompts.

Memory Module : Records agent behavior to support future decisions. Memory structures can be unified (short‑term only) or hybrid (short‑ and long‑term). Memory forms include language, database, vector representations, and lists. Operations cover reading, writing, and reflection.

Planning Module : Two categories— feedback‑free planning (single‑pass reasoning, multi‑path reasoning, external planners) and feedback‑driven planning (environment, human, or model feedback).

Action Module : Defines action goals (task execution, communication, exploration), generation methods (memory‑based recall or plan execution), action space (tool sets or model knowledge), and impact on environment, internal state, and future actions.

Key Challenges

Enhancing Role‑Playing Ability : Define role‑behavior relationships and evolution mechanisms; evaluate via metrics and scenarios; improve through prompt engineering or fine‑tuning.

Designing Memory Mechanisms : Analyze unified vs. hybrid memory, evaluate via metrics and scenarios, and study evolution (updates, autonomous refinement).

Improving Reasoning/Planning : Decompose tasks into subtasks, determine optimal execution order, and integrate external feedback effectively.

Efficient Multi‑Agent Collaboration : Define distinct roles, design cooperation and debate mechanisms, and establish convergence criteria.

Future Directions

LLM agents are heading toward two major tracks:

Task‑Specific Agents (e.g., MetaGPT, ChatDev, Ghost, DESP) aiming for alignment with correct human values and superhuman capabilities.

Real‑World Simulation Agents (e.g., Generative Agent, Social Simulation, RecAgent) focusing on diverse value representation and ordinary‑human alignment.

Current pain points include:

Hallucination : Cumulative hallucinations across interaction steps; solutions involve efficient human‑AI collaboration frameworks and robust human intervention mechanisms.

Efficiency : High latency when agents invoke many APIs; performance tables illustrate time costs for varying API counts.

Case Studies

User‑Behavior Simulation Agent : Combines LLMs with user behavior analysis. It contains profile, memory (sensory, short‑term, long‑term), and action modules, enabling agents to browse recommendation systems, converse, and post on social media, revealing emergent social patterns.

Multi‑Agent Software Development : Different agents assume roles such as CEO, CTO, coder, tester, and document writer, collaborating through communication to develop a complete software product.

References

Lei Wang, Chen Ma, Xueyang Feng, et al., "A Survey on Large Language Model based Autonomous Agents," CoRR abs/2308.11432 (2023).

Lei Wang, Jingsen Zhang, Hao Yang, et al., "When Large Language Model based Agent Meets User Behavior Analysis: A Novel User Simulation Paradigm."