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

The rapid maturation of large language models (LLMs) has led to the rise of LLM‑based AI agents, which are increasingly influencing various application domains. This article first introduces the five main topics that will be covered: the overall architecture of LLM‑based agents, key and difficult problems, user‑behavior simulation agents, multi‑agent software development, and future research directions.

1. Overall Architecture – An LLM‑based agent consists of four primary modules: a profiling module that describes background information (demographics, personality, social data) and can be built via manual prompt design, LLM‑generated prompts, or data‑alignment methods; a memory module that records actions and supports future decisions, with unified or hybrid memory structures and various storage forms (language, database, vector, list); a planning module that may operate with or without external feedback, using single‑path, multi‑path, or external planners; and an action module that defines goals, generates actions (e.g., recommendation‑system clicks, dialogues, social‑media posts), specifies the action space, and evaluates the impact on environment and internal state.

2. Key and Difficult Problems – The article identifies two major challenges: hallucination, which accumulates across interaction steps and requires efficient human‑machine collaboration and intervention mechanisms; and efficiency, where the runtime cost grows with the number of API calls, as illustrated by a benchmark table.

3. User‑Behavior Simulation Agent – A concrete case study demonstrates an agent that simulates user behavior in recommendation systems. It includes a profiling module (attributes such as ID, name, age, interests), a memory module with sensory, short‑term, and long‑term memories, and an action module that can watch movies, navigate pages, converse with other agents, or post on social media. The simulation reveals emergent social phenomena and user behavior patterns.

4. Multi‑Agent Software Development – Another case study shows how multiple specialized agents (CEO, CTO, CPO, developers, testers, document writers) collaborate to build a complete software product, mirroring a software company’s workflow and enabling coordinated development through communication and updates.

5. Future Directions – LLM‑based agents are expected to diverge into two major tracks: (a) task‑oriented agents that aim to align with correct human values and surpass human capabilities (e.g., MetaGPT, ChatDev); and (b) reality‑simulation agents that model diverse values and emulate real‑world environments (e.g., Generative Agent, Social Simulation). The article also stresses the need to address hallucination and efficiency issues to advance the field.

References to recent surveys and papers are provided for readers seeking deeper technical details.