> For the complete documentation index, see [llms.txt](https://llm-agents-a-survey.gitbook.io/llm-agents-a-survey-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://llm-agents-a-survey.gitbook.io/llm-agents-a-survey-docs/start-here/readme.md). # LLM Agents: A Survey The English edition of an online survey that organizes some 500 papers by function to explain what LLM agents are and how they're built. LLMs don't stop at answering questions anymore. They read GitHub issues, fix the code and open a PR, drive a browser themselves to make a booking, and even design and run experiments. All of that happened in just 3 years. The trouble is that close to 1,000 papers a month now come out in this field. It's hard to get a sense of where to start. This book was written to organize that literature. {% hint style="info" %} This book is the online edition of the survey paper [LLM Agents: A Survey](https://doi.org/10.20944/preprints202608.0265.v1). The paper is built to be cited more than read; this book aims for the opposite. {% endhint %} ## What makes this book different? It's organized by **function**, not by framework. Instead of what LangChain does, it covers the problems any agent has to solve and the answers that have been proposed so far. Frameworks change within a year, but planning, memory, tool use, evaluation and safety remain problems. ![How this survey is organized](https://2294612168-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FVnLYABSAQm1JBEDnSJCD%2Fuploads%2Fgit-blob-95095403b5091d8fa86925fcd1bc258011ae30b5%2Ftaxonomy-overview.svg?alt=media) We paid particular attention to three things: * **It covers what doesn't work, too**: results showing that self-correction, multi-agent debate and elaborate scaffolds fall short of expectations get the same weight as the success stories. * **It looks at cost alongside accuracy**: a reported accuracy gain means something only when you see the tokens, latency and money that came with it. * **It's current**: it includes topics that only became established after 2024, such as reasoning models, computer-use agents and agent security. ## Where should you start? If you're new to the area, we'd recommend reading in order from [**Foundations and Taxonomy**](/llm-agents-a-survey-docs/survey/foundations.md). It explains why an agent has to be a loop and then goes step by step through a single turn, which makes the later chapters much easier to follow. If you need a particular topic right away, feel free to jump straight there: | What you want to know | Where to read | | ------------------------------------------------------------ | ----------------------------------------------------------------------------------------- | | How does an agent decide its next action? | [Planning and Reasoning](/llm-agents-a-survey-docs/survey/planning-reasoning.md) | | On a long task, what does it remember and what does it drop? | [Memory](/llm-agents-a-survey-docs/survey/memory.md) | | How do tool calls actually go back and forth? | [Tool Use and Action Execution](/llm-agents-a-survey-docs/survey/tool-use.md) | | Do more agents really make things better? | [Multi-Agent Systems](/llm-agents-a-survey-docs/survey/multi-agent.md) | | How do web, GUI and embodied environments differ? | [Interactive Environments](/llm-agents-a-survey-docs/survey/interactive-environments.md) | | Where are agents used in the real world? | [Applications](/llm-agents-a-survey-docs/survey/applications.md) | | How do you check that you built it well? | [Evaluation and Benchmarks](/llm-agents-a-survey-docs/survey/evaluation.md) | | What can go wrong? | [Safety, Security, and Trustworthiness](/llm-agents-a-survey-docs/survey/safety.md) | | Which problems are still unsolved? | [Future Directions and Conclusion](/llm-agents-a-survey-docs/survey/future-directions.md) | If you'd rather follow the papers one at a time, the [reading list](/llm-agents-a-survey-docs/resources/reading-list.md) is the place to begin. ## We maintain the paper list too The papers this book covers are kept up to date on GitHub. New papers are added there first. * [**Awesome LLM Agent Papers**](https://github.com/js-lee-AI/awesome-llm-agent-papers): the paper list for this survey, organized by section with a one-line note on each entry. * [**Awesome Agent Loop Papers**](https://github.com/js-lee-AI/awesome-agent-loop-papers): the list for the follow-up survey *The Agent Loop*, which goes one level deeper into the loop itself. If you spot a paper that's missing, please let me know with a PR or an issue. With 1,000 papers a month, I'm bound to miss some on my own. The same goes for the translations. Most of the languages this book is translated into aren't my first language, so if a sentence reads oddly or a term is translated wrong, please tell me in an [issue](https://github.com/js-lee-AI/awesome-llm-agent-papers/issues). ## Citing this book If this book or the paper helped you, please cite it as follows. ```bibtex @article{lee2026llmagents, title = {LLM Agents: A Survey}, author = {Lee, Jungseob}, year = {2026}, month = {August}, journal = {Preprints}, publisher = {Preprints}, doi = {10.20944/preprints202608.0265.v1}, url = {https://doi.org/10.20944/preprints202608.0265.v1} } ``` Other citation formats are collected on [Cite the Paper](/llm-agents-a-survey-docs/resources/cite-the-paper.md). --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://llm-agents-a-survey.gitbook.io/llm-agents-a-survey-docs/start-here/readme.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.