LLM Agent: A Survey on Methodology, applications and challenges

📝 Paper Summary

LLM Agent Taxonomy Multi-Agent Collaboration Agent Evolution

This survey proposes a unified taxonomy for LLM agents centered on methodology, analyzing how agents are constructed, how they collaborate in groups, and how they evolve over time.

Core Problem

Existing research on LLM agents is fragmented, with prior surveys focusing narrowly on specific applications or environments without connecting agent design principles to emergent behaviors.

Why it matters:

The rapid proliferation of agent papers makes it difficult for researchers to grasp the fundamental methodological connections between isolated systems.
Prior surveys often separate individual agent design from multi-agent collaboration, missing the architectural continuity between them.
Understanding the full lifecycle—from construction to evolution—is critical as agents transition from theoretical constructs to commercial systems like DeepResearch.

Concrete Example: A traditional survey might categorize agents solely by application (e.g., 'gaming agents' vs. 'coding agents'). This fails to reveal that a coding agent (like ChatDev) and a research agent (like DeepResearch) share identical underlying 'Centralized Control' collaboration architectures, obscuring reusable design patterns.

Key Novelty

Build-Collaborate-Evolve Framework

Deconstructs agents into a lifecycle taxonomy: 'Construction' (internal modules), 'Collaboration' (interaction topologies), and 'Evolution' (improvement mechanisms).
Unifies individual and multi-agent perspectives by showing how internal modules (like memory) support collective behaviors (like decentralized cooperation).

Architecture

The organizational framework of the survey, visualizing the 'Build-Collaborate-Evolve' taxonomy.

Evaluation Highlights

Provides a structured taxonomy covering 3 main dimensions: Construction, Collaboration, and Evolution.
Categorizes collaboration into 3 distinct architectures: Centralized Control, Decentralized Cooperation, and Hybrid Architectures.
Identifies 3 pathways for evolution: Autonomous optimization, Multi-agent co-evolution, and External resource integration.

Breakthrough Assessment

8/10

A comprehensive, high-utility survey that provides a clean, methodology-centered taxonomy for a chaotic field. While not introducing a new algorithm, its structural framework significantly aids understanding.

⚙️ Technical Details

Problem Definition

Setting: Systematic literature review and taxonomy construction for Large Language Model (LLM) Agent systems.

Inputs: Existing academic literature and commercial systems related to LLM agents.

Outputs: A unified methodological taxonomy classifying agents by Construction, Collaboration, and Evolution.

Pipeline Flow

Construction (Profile, Memory, Planning, Action)
Collaboration (Centralized, Decentralized, Hybrid)
Evolution (Self-learning, Co-evolution, Resource-driven)

System Modules

Agent Construction

Define the individual agent's internal architecture

Model or implementation: Various (e.g., Llama, GPT-4)

Agent Collaboration

Orchestrate interaction between multiple agents

Model or implementation: N/A (Architectural Pattern)

Agent Evolution

Improve agent performance over time

Model or implementation: N/A (Learning Mechanism)

Novel Architectural Elements

The 'Build-Collaborate-Evolve' taxonomy itself is the structural contribution, unifying fragmented subfields.
Distinction between 'Batch-Generated Dynamic Profiles' (stochastic population generation) and 'Human-Curated Static Profiles' (fixed roles).

Comparison to Prior Work

vs. Xi et al. (The Rise and Potential...): This paper adds the 'Evolution' dimension explicitly, whereas prior general surveys focused mostly on construction and application.
vs. Wang et al. (Survey on LLM-based Autonomous Agents): This paper introduces a more granular classification of collaboration (Centralized/Decentralized/Hybrid) and specific evolution pathways.
vs. Gao et al. (Retrieval-Augmented Generation for Large Language Models) [not cited in paper]: This survey treats RAG as a sub-component of Memory/Evolution, framing it within the broader agent lifecycle rather than as a standalone technique.

Limitations

As a survey, it does not propose a new executable model or benchmark.
The rapid pace of the field means the specific SOTA agents mentioned (e.g., DeepResearch) may be superseded quickly.
Does not provide quantitative experimental comparisons between the surveyed methods, only qualitative taxonomy.

Reproducibility

Code: https://github.com/luo-junyu/Awesome-Agent-Papers

The paper is a survey; the 'reproducibility' refers to the availability of the surveyed papers list, which is hosted on GitHub.

📊 Experiments & Results

Evaluation Setup

Qualitative analysis and classification of existing literature.

Metrics:

Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Agent Construction is shifting from static, hand-crafted profiles to dynamic, batch-generated profiles that simulate social behaviors.
Memory is evolving from simple text history to structured 'knowledge as memory' using advanced RAG and tool synthesis.
Collaboration is moving towards hybrid and dynamic architectures where the topology itself (who talks to whom) is optimized during the task (e.g., DyLAN).
Evolution is the next frontier, with agents beginning to update their own parameters, prompts, or toolkits autonomously through self-reflection and multi-agent debate.

📚 Prerequisite Knowledge

Prerequisites

Basic understanding of Large Language Models (LLMs)
Familiarity with agentic concepts (planning, memory, tools)
Knowledge of basic multi-agent system topologies

Key Terms

RAG: Retrieval-Augmented Generation—enhancing model responses by fetching relevant data from external sources.

CoT: Chain-of-Thought—a prompting strategy where the model generates intermediate reasoning steps.

ToT: Tree-of-Thought—a planning method allowing models to explore multiple reasoning paths and backtrack.

ReAct: Reasoning and Acting—a paradigm where agents generate reasoning traces and task-specific actions in an interleaved manner.

SFT: Supervised Fine-Tuning—training a model on labeled examples to adapt it to specific tasks.

RL: Reinforcement Learning—training agents to maximize a reward signal through trial and error.

DyLAN: Dynamic LLM-Agent Network—a framework that dynamically adjusts collaboration structures based on agent contribution.

MetaGPT: A multi-agent framework that encodes Standard Operating Procedures (SOPs) into agent roles for software development.

DSPy: Declarative Self-improving Python—a framework for programming LLMs that can optimize prompts and weights.

PE: Prompt Engineering—optimizing the textual input to an LLM to guide its behavior.