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

As large language models (LLMs) mature, LLM‑based AI agents are becoming increasingly prominent. This article provides a comprehensive overview of LLM‑based agents, covering their overall architecture, key technical challenges, representative applications, and future research directions.

Overall Architecture

The architecture of an LLM‑based agent can be divided into four main modules:

Profiling Module : Describes the background information of the agent, including demographic, personality, and social attributes. Three generation strategies are discussed: hand‑crafted prompts, large‑model generation from a few examples, and data‑alignment prompts.

Memory Module : Records agent behavior to support future decisions. It includes unified (short‑term only) and hybrid (short‑term + long‑term) memory structures, four memory forms (language, database, vector, list), and three operations (read, write, reflect).

Planning Module : Handles reasoning and action planning. Two categories are presented: planning without feedback (single‑path, multi‑path, external planner) and planning with feedback (environment, human, model feedback).

Action Module : Defines the agent’s possible actions, goals, generation methods, action space, and impact on the environment and future actions.

Key Challenges

Role‑Playing Ability : Defining and evaluating the agent’s capacity to assume roles, with metrics and scenarios for assessment, and improvement methods via prompt engineering or fine‑tuning.

Memory Mechanism Design : Exploring vector‑retrieval‑based and LLM‑summarization‑based memories, evaluation criteria, and evolutionary aspects such as autonomous updates.

Reasoning/Planning Ability : Enhancing task decomposition, optimal execution order, and integrating external feedback into the reasoning loop.

Multi‑Agent Collaboration : Designing role definitions, cooperation mechanisms, debate protocols, and convergence conditions for efficient teamwork.

Representative Case Studies

User Behavior Simulation Agent : A three‑module agent (profiling, memory, action) that simulates user interactions in recommendation systems, social media, and dialogues, revealing social phenomena through multi‑round simulations.

Multi‑Agent Software Development : Agents assume distinct software‑development roles (CEO, CTO, coder, tester, writer) and collaborate via communication to produce a complete software product.

Future Directions

LLM‑based agents can be categorized into two major directions: (1) solving specific tasks (e.g., MetaGPT, ChatDev) with alignment to correct human values and superhuman capabilities; (2) simulating real‑world environments (e.g., generative agents, social simulation) that reflect diverse human values.

Current pain points include hallucination—where errors accumulate across interaction steps—and efficiency concerns, especially as the number of API calls grows. Proposed mitigations involve designing effective human‑machine collaboration frameworks and intervention mechanisms.

For further reading, see the survey "A Survey on Large Language Model based Autonomous Agents" (Wang et al., 2023) and the user‑simulation paper "When Large Language Model based Agent Meets User Behavior Analysis" (Wang et al., 2023).