A Survey on Large Language Model based Autonomous Agents
Wang et al., 2023. arXiv:2308.11432. A reading guide to the most-cited survey of LLM-based agents: what its framework contains, what it covers, and how it maps onto the patterns documented elsewhere on this site.
This is a reference summary, not the paper. Read the original at arXiv:2308.11432.
“A Survey on Large Language Model based Autonomous Agents” (Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, and colleagues at Renmin University of China) reviews the field of LLM-based agents along three axes: how an agent is constructed, how agents are applied, and how they are evaluated. Its central contribution is a unified construction framework of four modules, Profile, Memory, Planning, and Action, that the authors argue encompasses most prior work [1].
The construction framework
The survey’s most reused idea is its four-module decomposition of an agent’s architecture. It is a vocabulary, not a runtime: most production frameworks implement some subset of these modules under different names, and reading the survey makes those names legible.
Defines who the agent is: its role, identity, and persona. The survey distinguishes handcrafted profile descriptions from LLM-generated personas, and notes that profiling choices shape downstream planning and action behaviour.
Stores and retrieves information across steps. The survey organises memory by structure (unified vs hybrid), by format, and by operation (reading, writing, and reflection), framing it as the mechanism by which an agent accumulates experience.
Produces the sequence of actions toward a goal. The survey splits planning into two families: planning without feedback (single-pass decomposition) and planning with feedback (iterative refinement informed by the environment, the model itself, or a human).
Carries decisions into effect. The survey covers action target, strategy, and the action space, including whether the agent acts through external tools or from internal knowledge, and whether action capability is augmented by fine-tuning.
Applications
The second part of the survey catalogues where LLM-based agents have been applied, grouped into three broad fields:
- Social science - social simulation, psychology, political science and economy, and jurisprudence, where agents stand in for human actors in simulated settings.
- Natural science - documentation and data management, experiment assistance, and science education.
- Engineering - civil and industrial automation, robotics and embodied AI, computer science and software development, and aerospace and other domains.
Evaluation
The third part surveys how the field measures agents, split into two strategy families:
- Subjective evaluation - human annotation and Turing-test-style judgement, used where behaviour is hard to score automatically.
- Objective evaluation - metric-based scoring against environments, multi-task benchmark suites, and software-style testing protocols.
The survey is explicit that no single benchmark captures agent quality, and that subjective and objective methods are complementary rather than interchangeable.
Why it is worth reading
Two reasons. First, the four-module framework is a durable mental model: once you can name the Profile, Memory, Planning, and Action concerns of any agent, framework documentation reads faster because you know which module each feature belongs to. Second, the survey is a citation map. Its reference list, kept current in the companion GitHub repository, is one of the better entry points into the primary literature for a practitioner starting from zero.
How it maps onto this site
The survey is architecture-first; this site is pattern-first. They line up cleanly. The survey’s Planning module is where the ReAct and plan-and-execute patterns live, and its planning-with-feedback family is exactly the reflection pattern. Its Action module is the subject of the tool-use reference, and its Memory module is treated in the memory and RAG tooling page. Reading the survey for the framework and this site for the patterns covers both halves.
Citation and versions
The paper has been revised repeatedly since first posting. The arXiv submission history runs from v1 (22 August 2023) to v7 (2 March 2025), the latest a 35-page revision correcting typos. When citing it, cite the version you read; v1 and v7 differ in coverage. The canonical reference:
L. Wang, C. Ma, X. Feng, Z. Zhang, et al. “A Survey on Large Language Model based Autonomous Agents.” arXiv:2308.11432 (2023).
Frequently asked questions
What is arXiv:2308.11432?
It is the arXiv identifier for “A Survey on Large Language Model based Autonomous Agents” by Lei Wang and colleagues, first posted on 22 August 2023. It is one of the most-cited survey papers on LLM-based agents.
What framework does the survey propose?
It proposes a unified architecture for constructing an LLM-based agent from four modules: Profile, Memory, Planning, and Action. The authors argue this framework subsumes the majority of prior agent-construction work.
Is the survey still being updated?
Yes. The arXiv record shows revisions from v1 (August 2023) through v7 (March 2025), and the authors maintain a companion GitHub repository of references so the literature list stays current beyond the frozen PDF.
How does this survey differ from the Xi et al. survey?
The Wang et al. survey (arXiv:2308.11432) centres on a construction framework plus application and evaluation. The Xi et al. survey, “The Rise and Potential of LLM-Based Agents” (arXiv:2309.07864), is a companion published weeks later that leans toward capabilities, society-of-agents, and risk. They are complementary, not duplicative.
- Further Reading - the full working bibliography
- Patterns Index - the architectural patterns the survey’s Planning and Action modules cover
- How AI Agents Work - the perceive-reason-act-evaluate loop
Sources and Further Reading
- L. Wang, C. Ma, X. Feng, Z. Zhang, et al., A Survey on Large Language Model based Autonomous Agents, arXiv:2308.11432 (2023, rev. 2025).
- Companion reference repository, Paitesanshi/LLM-Agent-Survey (GitHub).
- Z. Xi et al., The Rise and Potential of LLM-Based Agents: A Survey, arXiv:2309.07864 (2023).