masamasa59/ai-agent-papers: A Curated, Four-Layer Reading List for LLM Agent Research
A collection of AI Agents papers (Updated biweekly)
At a glance
- What is it?
- The repository is a hand-picked arXiv reading list for AI agent research, organized into capabilities, architecture, operations and applications, with monthly trend newsletters. It is a reading aid, not a library you install, and its value depends on trusting the maintainer's selection criteria.
- Who is it for?
- Adopt this list if you need a filtered entry point into agent research and accept one maintainer's judgement about what is novel. Skip it if you need exhaustive coverage, reproducible literature search, or a citable survey with a stated methodology.
- 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 17 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the Repository Actually Is, and Who It Is For
This is a curated bibliography, not a framework, not a benchmark harness, and not runnable code in any meaningful sense. The README states the intent directly: the maintainer performs weekly arXiv searches using specific keywords and picks only papers that introduce a distinctively new approach or novel concept. The README also says the project does not strive for comprehensiveness. That single sentence defines the audience. If you are an engineer or researcher trying to keep up with agent literature and you want someone else to do the first-pass filtering, this list is aimed at you. If you need a complete census of everything published on LLM agents in a given month, it is explicitly not that.
The primary language field says Python, but the Python here is tooling around the list, not the list itself. The visible script is scripts/update_readme_badges.py, which regenerates the recent-addition badges in the README. There are no releases, no homepage, and the licence is not stated in the supplied material. Treat the repository as a document collection with a small maintenance script attached.
The Four-Layer Taxonomy and How Papers Get Filed
The organizing idea is that every paper lands in one of four layers: capabilities (what an agent can do), architecture (how it is built), operations (how it is run), and applications (where it is used). Each layer breaks into subclusters, and each subcluster links to a separate markdown file that holds the actual date-ordered entries. So capabilities splits into Core Cognition, Knowledge & Context, Action, Adaptation & Self-Improvement, Trust & Measurement, and Other. Architecture splits into Agent Design & Frameworks, Multi-Agent Systems, and Harness. Operations covers AgentOps & Observability, Human-AI Interaction & UX, and Governance. Applications splits along three axes at once: by interface (embodied, computer-use, web, mobile), by domain (finance, enterprise, AI scientist, vertical), and by system pattern (coding, data, deep research, world simulation).
The README points to TAXONOMY.md for the full directory map and the rules for where each paper is filed. That file is the part worth reading before you trust the structure, because the hard question for any taxonomy is what happens at the boundaries. A paper about a coding agent that also introduces a new memory mechanism could plausibly sit under applications/system/coding-agents.md or capabilities/knowledge-context/memory.md. The README does not resolve that in the text shown; TAXONOMY.md is where the tie-breaking rule would live. If it is vague there, expect duplicates or inconsistent placement over time.
Badges, Cluster Counts and the Regeneration Script
The README carries inline counts next to cluster headings, formatted as (+N) for additions in the last two months, with a flame marker indicating high activity. A blockquote explains the convention: badges show papers added in the last 2 months, cluster headings show the sum, and the example given is Jul to Aug 2026. The same blockquote gives the regeneration command verbatim: python scripts/update_readme_badges.py.
This is the one piece of the project with a mechanical contract. The counts are derived, not hand-typed, which means they can drift if the script is not re-run after papers are added. There is no CI configuration described in the supplied material, so nothing forces the badges to stay current. If you fork the repository and add papers, run the script or your counts will be wrong. That is a small operational detail, but it is the difference between a list that stays self-consistent and one that slowly accumulates stale numbers.
Monthly Trend Newsletters and the Reading Procedure Behind Them
Below the main index, the repository hosts trend newsletters organized by month, from 2026-04 through 2026-08. The README describes the method for issues from 2026-06 onward: the maintainer reads the arXiv HTML full text of each paper, cites figures, and centers the write-up on facts corroborated across multiple papers. The stated procedure lives in .claude/skills/newsletter, which suggests the workflow is partly automated through a Claude skill definition rather than being entirely manual.
The August 2026 set covers Harness, Safety, Agent Evaluation, Self-Evolution, Skills, Failure Attribution, Agent Tuning, and Governance. July 2026 covers Harness, Agent Evaluation, and Self-Evolution. June 2026 covers Self-Evolution, Coding Agents, and Skills. May 2026 is a single general trends file, and April 2026 covers Self-Evolution, Memory, and Tool Use. The asymmetry is informative: Harness, Self-Evolution, and Agent Evaluation recur across multiple months, while others appear once. That recurrence is a signal about where the maintainer sees sustained activity, and it is more useful than any single count.
One caveat about the newsletters: the README says they are written in Japanese for the section headed 研究トレンド. If your team reads English only, the newsletters are not usable without translation, even though the main index and the paper entries are in English.
Where This List Breaks Down
The selection rule is the main limitation. Papers are added when they stand out from existing methods, judged by one person doing keyword searches. That is a reasonable editorial policy and a poor discovery mechanism. Keyword-based arXiv search will miss work that uses different terminology for the same idea, and a novelty judgement made at first read will occasionally be wrong in both directions: some included papers will look thin a year later, and some excluded ones will turn out to matter. There is no stated inclusion checklist, no second reviewer, and no appeal process.
The licence is not stated in the supplied material. That matters more than it does for a code project. A repository that aggregates links and short descriptions is in different territory from one that reproduces figures from arXiv papers, which the newsletter description says it does. arXiv papers carry their own licences, and those vary. If you plan to reuse newsletter figures or text in your own internal documents, verify the licence of each source paper and the licence of this repository before doing so. This is not legal advice, just a reason to check before copying.
Finally, there is no versioning story. No releases were retrieved, and the default branch is main. If you cite this list in a report, cite the commit hash, because the content moves.
How It Compares to the Lists It Cites
The README's References section names three alternatives: zjunlp/LLMAgentPapers, hyp1231/awesome-llm-powered-agent, and kaushikb11/awesome-llm-agents. The difference in approach is the interesting part. Awesome-style lists generally optimize for breadth and accept contributions, which produces long indexes with uneven annotation and no consistent filing rule. This repository does the opposite: it restricts intake by a novelty filter and imposes a fixed four-layer taxonomy with a separate TAXONOMY.md defining placement. The trade is coverage for navigability.
That trade is not obviously correct. If you are starting a literature review, an awesome list with more entries gives you a wider net to filter yourself. If you already know roughly what you are looking for and want to avoid reading forty abstracts to find three relevant papers, the tighter list wins. The newsletters are the part the alternatives do not replicate: a monthly synthesis that reads full texts and cross-references papers is a different product from a link index, and it is the strongest reason to bookmark this repository rather than one of the others.
Maintenance Cost and Upgrade Path
For a consumer of the list, there is nothing to upgrade. You read markdown files and follow arXiv links. The cost is attention, not installation.
For anyone who wants to contribute or fork, the maintenance surface is small but real. Adding a paper means editing the correct cluster file, choosing the right layer, and then running python scripts/update_readme_badges.py so the (+N) counts and flame markers stay accurate. The badge logic assumes a two-month window, so the script needs to be re-run periodically even when no new papers are added, or the counts will overstate recent activity. Newsletter production is heavier: the README describes reading full arXiv HTML, citing figures, and corroborating facts across papers, with the procedure documented under .claude/skills/newsletter. That is a per-issue editorial commitment, not a scripted pipeline, and it is the part most likely to lapse if the maintainer steps back.
The licence question also affects forking. With no licence stated in the supplied material, the default position under most copyright regimes is that no rights are granted beyond what the platform's terms allow. If you intend to redistribute the newsletters or embed them in a product, resolve that first.
Editorial conclusion
Adopt this list if you need a filtered entry point into agent research and accept one maintainer's judgement about what is novel. Skip it if you need exhaustive coverage, reproducible literature search, or a citable survey with a stated methodology. Before relying on it, open TAXONOMY.md to check the filing rules, confirm the licence situation, and read one newsletter such as newsletters/aug_2026/harness_trends.md to see whether the level of detail matches what your team needs.
Community notes