Overview of Large Language Model‑Based AI Agents: Architecture, Challenges, and Future Directions

Introduction With the rapid maturation of large language models (LLMs), AI agents built on these models are becoming increasingly visible. This article surveys the core concepts of LLM‑based agents and explores important application directions in the era of large models.

Five Main Topics

Overall Architecture An LLM‑based agent consists of four primary modules: profile, memory, planning, and action.

Key and Difficult Issues Challenges include enhancing role‑playing ability, designing effective memory mechanisms, improving reasoning/planning, and enabling efficient multi‑agent collaboration.

User‑Behavior Simulation Agent Demonstrates how an agent can simulate user actions in recommendation systems, social media, and dialogue, using a profile module (demographics, personality, social info), a memory module (short‑term, long‑term, reflective), and an action module (environment interaction).

Multi‑Agent Software Development Describes a scenario where multiple specialized agents (CEO, CTO, coder, tester, documenter) cooperate to develop a complete software product, mirroring a software company workflow.

Future Directions Two major trends are identified: (a) task‑oriented agents that align with correct human values and surpass human capabilities, and (b) agents that simulate real‑world environments, emphasizing diverse value alignment and human‑like behavior.

Detailed Module Descriptions

Profile Module Captures background information through demographics, personality, and social data. Generation strategies include manual prompt design, large‑model generation from few examples, and data‑alignment methods.

Memory Module Records agent behavior to support future decisions. It can be a unified short‑term memory or a hybrid of short‑ and long‑term memory, stored as language, database entries, vector representations, or lists, and supports read, write, and reflection operations.

Planning Module Includes non‑feedback planning (single‑pass, multi‑path, external planner) and feedback‑driven planning that incorporates environment, human, or model feedback.

Action Module Defines action goals (task completion, communication, exploration), generation methods (memory‑driven or plan‑driven), action spaces (tool sets or model knowledge), and impacts on environment and future actions.

Challenges

Hallucination Cumulative errors arise from repeated interactions; solutions involve efficient human‑machine collaboration frameworks and robust human intervention mechanisms.

Efficiency Performance bottlenecks appear as API call volume grows; empirical data on latency across different agents is provided.

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.”

Additional related works on MetaGPT, ChatDev, Ghost, DESP, Generative Agent, Social Simulation, and RecAgent are mentioned.