Large Language Model Agent: A Survey on Methodology, Applications and Challenges
TL;DR Highlight
A comprehensive survey organizing the latest LLM agent research into a 3-axis framework: building, collaborating, and evolving.
Who Should Read
AI engineers and architects building or evaluating LLM agent systems. Developers who want a one-stop overview of multi-agent design patterns, security threats, and real-world applications.
Core Mechanics
- Agent architecture decomposes into 4 stages: Profile Definition → Memory → Planning → Action Execution, forming a cyclic optimization loop at each layer
- Multi-agent collaboration patterns fall into 3 categories: Centralized (controller directs), Decentralized (agents communicate directly), and Hybrid (mix of both)
- Agent evolution also has 3 approaches: self-evolution through experience, environment-driven adaptation, and human-guided refinement
- Security threats are categorized across the full pipeline: prompt injection, memory poisoning, external source poisoning, and action-level attacks
Evidence
- AgentBench tests agents across 8 interactive environments; Mind2Web evaluates on 137 real websites across 31 domains
- AgentHarm benchmark collected 440 malicious agent tasks across 11 risk categories — the first systematic evaluation of LLM agent misuse
- OSWorld tests 369 multi-app tasks across Ubuntu/Windows/macOS for cross-platform agent evaluation
How to Apply
- When designing a new agent system, use this paper's 4-pillar checklist (Profile/Memory/Planning/Action) as a template. Pick implementations (ReAct, MemGPT, ToT, etc.) for each component to structure your design decisions.
- If running a RAG-based agent, audit for External Source Poisoning threats — adversaries can inject malicious content into retrievable documents to manipulate agent behavior.
Code Example
# LLM Agent 4-pillar design checklist template
agent_design = {
# 1. Profile Definition
"profile": {
"type": "static", # or 'dynamic'
"role": "software engineer",
"system_prompt": "You are an expert Python developer..."
},
# 2. Memory Mechanism
"memory": {
"short_term": "conversation history (last N turns)",
"long_term": "vector DB (e.g., ChromaDB, Pinecone)",
"retrieval": "RAG with top-k=5"
},
# 3. Planning Capability
"planning": {
"strategy": "ReAct", # or 'ToT', 'Plan-and-Solve'
"decomposition": "single-path chain", # or 'tree'
"feedback": ["environment", "self-reflection"]
},
# 4. Action Execution
"action": {
"tools": ["search", "code_interpreter", "calculator"],
"tool_selection": "LLM-based",
"physical": False # True for embodied agents
}
}
# Centralized vs Decentralized selection guide
collaboration_patterns = {
"strict_workflow": "Centralized (MetaGPT pattern) - clear role division",
"creative_debate": "Decentralized (MAD pattern) - multiple agents debate",
"adaptive": "Hybrid (MDAgents) - automatic routing based on complexity"
}Terminology
Related Resources
Original Abstract (Expand)
The era of intelligent agents is upon us, driven by revolutionary advancements in large language models. Large Language Model (LLM) agents, with goal-driven behaviors and dynamic adaptation capabilities, potentially represent a critical pathway toward artificial general intelligence. This survey systematically deconstructs LLM agent systems through a methodology-centered taxonomy, linking architectural foundations, collaboration mechanisms, and evolutionary pathways. We unify fragmented research threads by revealing fundamental connections between agent design principles and their emergent behaviors in complex environments. Our work provides a unified architectural perspective, examining how agents are constructed, how they collaborate, and how they evolve over time, while also addressing evaluation methodologies, tool applications, practical challenges, and diverse application domains. By surveying the latest developments in this rapidly evolving field, we offer researchers a structured taxonomy for understanding LLM agents and identify promising directions for future research. The collection is available at https://github.com/luo-junyu/Awesome-Agent-Papers.