LLMAgentPapers is a single README file with an MIT badge, a .DS_Store, and no LICENSE
Must-read Papers on LLM Agents.
At a glance
- What is it?
- zjunlp/LLMAgentPapers is a curated reading list for LLM agent research, and the repository contains exactly two files. The badge advertises MIT while nothing in the repository records a license, and the News section stopped in March 2024 while the branch kept taking pushes.
- Who is it for?
- Use this list as an index into the agent literature and treat the file itself as the artifact under review, because there is no other one. It is worth knowing that the entries are dated entries with no venue field, that the three newest overview items sit on two different preprint servers rather than arXiv, and that the table of contents at the top is maintained separately from the entries below it.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- zjunlp/LLMAgentPapers3,146 ★ · +32
| Repository | 2026-09-10 | 2026-11-15 |
|---|---|---|
| zjunlp/LLMAgentPapers | 3.1k | 3.1k |
32 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The whole repository is one README and a .DS_Store
The tree at the root of zjunlp/LLMAgentPapers contains two entries: `README.md` and `.DS_Store`.
There is no code, no data file, no generator script and no configuration. The project is a document, and the document is the entire deliverable. That is a legitimate shape for a reading list, and it also means every property that matters is a property of one markdown file that any number of people can edit at once.
The `.DS_Store` file is the odd one. It is the metadata macOS Finder writes into a directory, it serves no function in a repository, and its presence at the root tells you the file was committed from a Mac file browser rather than added deliberately. It is harmless and it is also the only thing in the tree that no one chose to put there.
The repository records no primary language and no homepage, which is consistent with a repository containing no source files to classify.
The badge says MIT, nothing records a license, and there is no license file
Near the top of the file there is a badge linking to the MIT license text at opensource.org. The repository itself records no license, and the tree has no LICENSE file to open.
So the three sources disagree in a specific way. The badge asserts a license, the repository metadata stays silent, and the license text is not present in the repository at all. Somebody reading the badge gets a conclusion they cannot verify from the project.
This matters more than it might for a list of references. A paper list is the kind of thing people copy: a reading guide for new students, a reading list for a lab, a syllabus appendix. Copying the file copies the badge and with it an unbacked license claim.
The file does ask readers to look elsewhere for substance. It points at two sibling collections from the same owner, Prompt4ReasoningPapers for reasoning with language model prompting and KnowledgeEditingPapers for knowledge editing in large language models, presented in a quoted block rather than as a plain section.
The News section stops in March 2024 while the branch was pushed in October 2026
There is a News section with two entries and it is the clearest signal of how this list is maintained.
The first entry is dated 2023-06 and says the repository was created to maintain a paper list on multi-agents. The second is dated 2024-03 and announces one new paper, KnowAgent: Knowledge-Augmented Planning for LLM-Based Agents.
That is the most recent news item. The last push to the main branch was on 2026-10-03, two and a half years later.
So the file has been edited extensively since the last news entry without the news log moving. The entries below it do carry newer material, including overview items stamped 2026.3 and 2026.8, which means the paper list itself is current and only the announcement log is not.
The practical reading: trust the dated entries in the body, and do not use the News section to judge how current the list is. It answers that question for 2024 and no later.
It was created for multi-agent papers and now answers to agents in general
The title, the summary line, and the repository description all say LLM agents. The first news entry says something narrower: the repository was created in June 2023 to maintain a paper list on multi-agents.
The structure shows where it went. The contents block has a single Agent section subdivided into Personality, Memory, Planning, Tool use, and RL training, and beside it a separate Multiple Agents section subdivided into Task-Oriented Communication, itself split into Collaborative Exchanges and Adversarial Interactions, plus a Casual and Open Conversations subsection.
So a single-agent and a multi-agent treatment sit at the same level, with the multi-agent half carrying its own internal taxonomy. That is a reasonable structure for a field where communication between agents became its own subfield, and it is also a reminder that the two halves have different densities of literature behind them.
The remaining contents entries are Application, Framework, and Others, then a Resources block holding Benchmarks, Types of Tools, and a Tool List.
Three of the twelve overview entries are not on arXiv
The Overview subsection carries twelve numbered surveys, each with an inline author list, a link, and a date stamp in the form year.month.
Nine of the twelve resolve to arxiv.org. The other three do not. Entry ten, Harness Engineering for Language Agents, points at a preprints.org manuscript. Entry eleven, LLM Agents: A Survey, points at another preprints.org manuscript. Entry twelve, The Agent Loop, points at an SSRN abstract page:
12. **The Agent Loop: A Survey of Control Strategies, Skills, and Harnesses for LLM Agents**
*Jungseob Lee, Chanjun Park.* [[abs](https://ssrn.com/abstract=7186738)], 2026.8Three preprint hosts appear in one list of twelve, which is a fair reflection of where this literature is being posted now. It also means the entries are not uniform: an SSRN abstract page and an arXiv abstract page are different kinds of object, and neither is peer review.
The link labels are not uniform either. Most read `abs`, but two of the older entries, Agent AI and Personal LLM Agents, carry an `abs` label pointing at a `.pdf` path rather than an abstract page.
Emoji-derived anchors carry stray characters, and one name appears on two entries
The contents block at the top of the file is built from emoji headings, and the anchors it generates are fragile. The Planning anchor ends with a variation selector appended after the hyphen, and the Framework anchor begins with one before the hyphen. Those characters end up inside the link target, which is exactly the sort of thing that survives in a file for years because nobody clicks every entry.
The Casual and Open Conversations entry is a second case. Its title contains a slash, and the anchor for it reads without one, so the link text and the target disagree about the name of the section.
A separate detail in the Overview is author overlap. Entry eleven is a single-author paper by Jungseob Lee, and entry twelve, carrying the same date stamp, adds Chanjun Park to the same name. Two of the three newest overview items, on two different preprint servers, come from the same author, which suggests the person maintaining this list is also publishing in the area.
That is worth knowing when you weigh a list against a survey of the field.
Author lists are inlined, and the longest carries thirty-three names
Each entry in the papers sections follows the same shape: a numbered title, then the author list in italics on its own line, then the link and the date stamp.
Inlining the authors means every entry carries its own copy of the bibliography rather than pointing at one. The cost is visible in the file. The Rise and Potential of Large Language Model Based Agents entry runs to thirty-three names, Personal LLM Agents to twenty-nine, and The Landscape of Agentic Reinforcement Learning for LLMs to twenty-seven. Interactive Natural Language Processing lists twenty-two.
Only one Overview entry has a single author, and two of the Personality entries list six or seven. So the list is not uniform in authorship, which is a property of the field rather than of the curation, and it means the file's own bulk is concentrated in a handful of mega-consortium surveys.
The date stamps are the other repeated element. They run from 2023.2 in the Personality subsection through 2023.5, 2023.4, 2023.5 in the entries visible here, and they follow the arXiv identifier rather than a publication venue, since no entry in the file names a conference or journal.
The contents block promises nine sections past the overview, and lists itself twice
The contents block at the top of the file runs to twenty-six entries: the title, then News and Content, then Papers with Overview beneath it, five Agent subsections, two Multiple Agents subsections with one nested pair, Application, Framework, and Others, then the Resources block of Benchmarks, Types of Tools, and a Tool List, and finally Contribution with its contributing guide and contributor list.
Two of those twenty-six are News and Content itself. A table of contents that indexes the news log and its own contents section is a file that has to be edited by hand in two places for every heading anybody adds, which is the likeliest explanation for the anchor problems elsewhere in it.
The body entries visible here stop partway through the first of those five Agent subsections, at the fifth reference in Personality, which is the Theory of Mind paper from 2023.2 followed by four entries stamped 2023.4 and 2023.5. Everything the contents block promises after that point has to be read further down the same file.
The Overview is the only heading in the block with no emoji in front of it. Every other heading carries one, which makes the first level of the hierarchy the only one a plain-text grep for a section name will find cleanly.
Editorial conclusion
Use this list as an index into the agent literature and treat the file itself as the artifact under review, because there is no other one. It is worth knowing that the entries are dated entries with no venue field, that the three newest overview items sit on two different preprint servers rather than arXiv, and that the table of contents at the top is maintained separately from the entries below it. If you plan to reuse any of it, the license question is unresolved: the badge points at the MIT license text, the repository records none, and there is no license file to read, so get that answered before you copy a list into something you publish.
Frequently asked questions
What is zjunlp/LLMAgentPapers?
It is a must-read paper list on large language model agents, organised into an overview of twelve surveys, an Agent section covering personality, memory, planning, tool use, and RL training, a Multiple Agents section, application and framework sections, and a resources block. The repository contains only a README file and a .DS_Store.
What license does LLMAgentPapers use?
The README carries a badge linking to the MIT license text, but the repository records no license and the tree has no LICENSE file. The badge is the only statement of terms in the project.
How current is the LLMAgentPapers list?
The News section's most recent entry is dated 2024-03, announcing the KnowAgent paper, while the main branch was last pushed on 2026-10-03. The body of the list is newer than the news log, carrying overview entries stamped 2026.3 and 2026.8.
Are all the papers in LLMAgentPapers hosted on arXiv?
No. Nine of the twelve overview entries point at arxiv.org. Two point at preprints.org manuscripts and one points at an SSRN abstract page, so three preprint hosts appear in a single list of twelve.
How does LLMAgentPapers ask for contributions?
The contents block includes a Contribution section with a contributing guide and a contributor list, and the file promotes two sibling collections from the same owner, Prompt4ReasoningPapers and KnowledgeEditingPapers. The repository has no GitHub releases.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/zjunlp-llmagentpapers)Star history chart for your README
[](https://hysenlabs.com/projects/zjunlp-llmagentpapers)Was this useful?