AGI-Papers: A Curated Korean-Language Paper Archive for Agents, RAG and On-Device AI

What AGI-Papers actually is, and who it is written for

The repository describes itself as "a curated archive of breakthroughs in Agents, Architecture, Training, RAG, and On-Device AI." That is the whole product. There is no package to install, no model to run, no API to call. What you get is a set of markdown files, each one a review of a paper, filed under a topic directory and listed in the README with a one-line Korean summary.

The intended reader is someone who already follows the field and wants a shortcut through it. The README says the reviews mostly track papers the author covered on LinkedIn, and that pre-release insights or rougher thoughts are written here first, before they go to social media. So the archive sits upstream of the author's public commentary. If you want the polished version, you go to LinkedIn. If you want the draft, you come here.

The audience is narrower than the topic list suggests. Every entry title and summary in the README is in Korean, and the linked files are markdown under Agents/, Architecture/, RAG/, Evaluation/, On_Device/, Post_Training/, Pre_Training/, Projects/ and Trends/. An English-only reader can use the archive as a bibliographic index (paper names, arXiv-style titles, dates) but cannot read the analysis without translating it. That is a real filter, not a caveat.

The nine top-level topic folders map to the contents list: Agents, Architecture, Pre-Training, Post-Training, Evaluation, RAG & Knowledge, On-Device AI, Projects, and Trends & Industry, plus a Recommended Resources section. The README's badge row labels the status Active, and the last push to the repository was on 2026-09-15.

How the archive is organised: numbered files, topic folders and two side channels

The mechanism is a flat numbering scheme inside topic directories. Agent entries run from Agents/5.md through Agents/132.md, with gaps. Architecture entries are numbered in the 140s. The number is an insertion order, not a version and not a rating. Nothing in the README explains what a gap means, so you should treat the numbering as a rough chronology rather than a complete sequence.

Two files sit outside the topic folders and change how you should read the rest. Drafts.md is described as "New Drafts (Working in Progress)", meaning entries there are unfinished. Pre-README.md holds "Past AGI-Papers (Pre-2026)", so older reviews are deliberately split off from the current index. If you only read README.md you will miss both the incomplete work and everything before 2026.

The PDF/ directory is the second side channel. The news section lists two lecture decks: Modern-Model-Architectures-Overview.pdf, updated 2026-03-20, and introduction-to-post-training-and-beyond.pdf, updated 2026-03-15. These are the only artefacts in the repository that are not markdown, and they are the closest thing to a structured teaching resource. The README does not say whether the decks correspond to specific entries.

There is also an image/ directory and a LICENSE file at the top level. The README does not document a table of contents generator, a link checker, or any script that validates the entries, so link rot in the numbered files is a maintenance risk the project has not addressed in what it publishes.

Reading your first AGI-Papers entry

There is no installer. The README gives no install command, no package name and no build step, because the repository is documentation. The only setup is getting a local copy so you can read the markdown offline and follow relative links between entries. The README does not document a clone command, so fetching the repository with git is the only route to a local copy.

After cloning, the top-level entries listed in the repository are .gitignore, Agents/, Architecture/, Drafts.md, Drafts/, Evaluation/, LICENSE, On_Device/, PDF/, Post_Training/, Pre-README.md, Pre_Training/, Projects/, RAG/, README.md, Trends/, and image/.

To start with the material the README leads with, open the first Agents entry. The README links it as ./Agents/132.md and summarises it as a multi-agent architecture that combines hundreds of parallel sub-agent readers with a central commander trained by GRPO, aimed at very long contexts. That file is plain markdown, so any text editor or markdown viewer opens it. What you should see is the entry body in Korean; the summary in the README is not repeated verbatim inside the file, so expect the file to carry the detail the one-line listing omits.

If you prefer the lecture format, the two decks sit under PDF/ and can be opened with any PDF reader. The README links them as ./PDF/Modern-Model-Architectures-Overview.pdf and ./PDF/introduction-to-post-training-and-beyond.pdf. There is no documented way to regenerate them from the markdown, so treat them as separate deliverables.

Where the archive breaks down: language, coverage gaps and unverifiable claims

The first limitation is the language barrier, and it is structural rather than cosmetic. The README is Korean. The entries are Korean. The two PDF filenames are English but the decks are described in Korean news items. A reader who does not read Korean gets a list of paper topics and nothing else. For a repository whose stated ambition is "Toward Artificial General Intelligence (AGI) in 2026", the reach is limited to one language community.

The second limitation is what the listings claim. The README summaries carry specific numbers: an 11x speed improvement, a 75% reduction in token waste, a 24x speed advantage, 896K-token contexts. These are the author's summaries of the papers, not results reproduced in this repository. There is no code, no evaluation harness and no reproducibility note anywhere in the top-level layout. If you need to cite a figure, cite the paper, not the archive.

The third limitation is coverage. The topic folders are uneven. Agents/ is by far the deepest, running past entry 132. Architecture/ entries sit in the 140s. Other folders are not enumerated in the README excerpt, so you cannot tell from the index alone how thin Evaluation/ or Projects/ is. The README also does not state a selection criterion beyond the author's own reading, which means the archive inherits one person's interests and blind spots.

Finally, the README does not document how entries are added, whether corrections are accepted, or what a reader should do when a linked paper is retracted or superseded. There is no changelog beyond the two news items about PDF updates.

AGI-Papers versus Papers with Code and Awesome lists

The obvious comparison is with Papers with Code, which indexes papers alongside implementations, datasets and benchmark leaderboards. The difference in approach is that Papers with Code is machine-assembled and code-centric: a paper appears with its repository, its results table and its task. AGI-Papers is human-assembled and prose-centric. An entry exists because the author read the paper and wrote about it. There is no leaderboard, no linked implementation, and no task taxonomy.

A second comparison is the Awesome-list pattern, where a README is the entire artefact and entries are bare links with a few words. AGI-Papers is closer to that, but each link points to an in-repository markdown file rather than to the paper. That is the meaningful design choice: the archive stores the review, not just the reference. It also means the archive's value depends entirely on the quality of reviews you cannot see from the index.

A third comparison is a personal blog. This is effectively that, with git history and folder structure attached. The advantage over a blog is that entries are addressable and can be linked to a specific number, and that the pre-2026 material is preserved in Pre-README.md rather than buried in pagination. The disadvantage is that a blog usually ships an RSS feed and a search box; the README does not mention either.

If your goal is to find the code for a paper, AGI-Papers is the wrong tool. If your goal is to find out which papers one experienced reader considered worth writing about, in a fixed topic structure, it does that job.

Licence, maintenance and the cost of keeping an archive current

The repository is MIT licensed. In practice this matters for two things: you can reuse the review text and the folder structure in your own project, and you can redistribute the markdown, provided the copyright notice and permission notice travel with it. It does not grant you rights to the papers being reviewed, which remain with their authors and publishers. The two PDFs under PDF/ are lecture slides by the repository author; the README does not state a separate licence for them, so the MIT file at the top level is the only licence signal available. This is a description of what the repository states, not legal advice.

Upgrade cost is close to zero in the software sense, because there is nothing to upgrade. The cost is editorial. Every entry is a manual write-up, and the numbering scheme means new entries do not renumber old ones, so links stay stable. The last push was on 2026-09-15, and the only dated news items are the two PDF updates from March 2026. The README does not describe a release process, and no releases were retrieved.

The real maintenance burden falls on the reader. Because entries are numbered and topic folders are uneven, there is no way to tell from the README whether a topic you care about has been touched recently. You have to open the folder and look at the highest number. That is a small friction, but it repeats every time you visit.

Editorial conclusion

Adopt AGI-Papers if you read Korean and want a dated, topic-sorted pointer list to recent agent and architecture work, with PDF slides for two topics. Do not adopt it if you need English summaries, executable code, or a stable citable index: entries are numbered markdown files whose contents the README only hints at, and the archive is a personal reading log rather than a peer-reviewed survey. Before relying on it, open Drafts.md to see what is unfinished, check Pre-README.md for the pre-2026 entries, and confirm that the two PDFs under PDF/ cover the ground you need.

Frequently asked questions

Official sources

1. gyunggyung/AGI-Papers on GitHub
2. Issues
3. License: MIT
4. Project website
5. README