LLM_MultiAgents_Survey_Papers: A Reading List, Not a Runtime

What the repository actually contains

The repository is a bibliography. Its README opens with a link to the survey paper at arXiv 2402.01680 and then presents three images: a trend chart, an architecture diagram labelled LLM-MA, and an overview table. Below those images sit the paper entries themselves, each one a date, a title, an author list and a link. Most links point to arXiv. A minority point to GitHub repositories, and the crewAI entry is the clearest example, listed under Multi-Agents Orchestration and Efficiency with a repo link rather than a paper link. The metadata retrieved for this review lists no primary language, no licence and no releases. That is consistent with what the README shows: there is nothing in it to compile, install or import. Anyone arriving here expecting a framework has misread the artefact. The correct expectation is a starting point for a literature review, and the value it adds is ordering rather than code.

The five streams and what each one is for

The README states that papers are categorised into five main streams: Multi-Agents Framework, Multi-Agents Orchestration and Efficiency, Multi-Agents for Problem Solving, Multi-Agents for World Simulation, and Multi-Agents Datasets and Benchmarks. The table of contents expands two of these. Problem Solving is split into Software development, Embodied Agents, Science Team for Experiment Operations, Science Debate and Database. World Simulation is split into Society, Game, Psychology, Economy, Recommender System, Policy Making and Disease propagation Simulation. This is the part of the repository with the most practical use. If you are writing a related-work section on multi-agent debate, the Orchestration and Efficiency stream is where the 2023/05 entry on encouraging divergent thinking through multi-agent debate sits. If you are building a simulated society, the World Simulation subsections give you the sub-area split before you start searching. The taxonomy is doing real work here: it is the difference between a keyword search and a structured reading plan.

How the list is maintained, and what the dates mean

The News section records two events. In 2024/01 the repository was created. In 2024/02 the maintainers stated they would update the paper list every two weeks and include the listed papers in the next version of the paper, with an invitation to contact them about anything missed. Each entry carries a bracketed date such as [2024/03] or [2023/08], which appears to be the date the entry was added or the paper posted, though the README does not define it explicitly. The retrieved metadata shows a last push of 2026-08-21, and the Orchestration and Efficiency section contains entries dated [2025/10] and [2026/04]. Those later dates are inconsistent with a repository that stopped in early 2024, which suggests the list has continued to receive additions well past the original announcement. Treat the two-week cadence as a stated intention rather than a guarantee: the README says it, but nothing in the repository enforces it, and there is no changelog to check against.

There is no install step, and that is the point

You cannot run this repository. There is no package manifest visible in the material, no setup instructions, no configuration keys and no command examples in the README. The only operational instruction is the Contributing section referenced in the table of contents, which the truncated README does not expand. So the honest answer to how you get it running is that you clone it and read it, or you read the rendered README on the hosting page. If you want to reproduce any result discussed in the survey, you follow the arXiv link for that paper and work from that paper's own repository. This repository will tell you which paper to look at. It will not tell you how to run it, what its dependencies are, or whether the code still works. For a survey companion that is a defensible scope. For anyone who skimmed the title and expected a multi-agent library, it is a dead end.

The licensing gap is the sharpest limitation

The metadata retrieved for this review lists no licence. The README contains a commented-out citation block, which means the citation text is present in the file but not rendered, and the repository's own licence terms are not stated anywhere in the material available. For a bibliography this matters less than it would for code, because you are not redistributing the repository's contents. It matters more if you intend to copy the taxonomy structure into your own documentation, or mirror the list. The papers themselves carry their own licences, and arXiv preprints vary. Before you build anything that depends on redistributing this material, check the repository for a licence file directly rather than assuming one exists because the project is public. The absence of a stated licence is a fact about the repository, not a legal conclusion, and this review does not offer one.

Where a different resource is the better choice

If your goal is to build or run a multi-agent system, the entries in this list are the alternatives, not the list itself. AutoGen, MetaGPT, CAMEL, AgentVerse and crewAI all appear as entries under the Framework and Orchestration streams, each with a link to its own paper or repository. Those are the projects with install instructions, configuration and runnable code. The difference in approach is straightforward: this repository organises the literature, and those projects implement it. Choosing between them is not a comparison of competing tools, because they do not compete. A reasonable workflow is to use this list to decide which framework matches your problem shape, then leave the list and go to that framework's own documentation. What this repository gives you that a framework's docs will not is the surrounding research context, including the critical papers. The Orchestration section lists an entry titled Achilles Heel of Distributed Multi-Agent Systems, which is the kind of counterweight a framework's own marketing will never surface.

Who should use it, and what to check first

Use it if you are scoping a literature review, preparing a related-work section, or trying to work out which sub-area of LLM multi-agent research your problem belongs to. The five-stream split and the Problem Solving and World Simulation subsections save time that a raw arXiv search would burn. Do not use it as a dependency, a template or a source of runnable examples, because it contains none. Two things to verify before you rely on it. First, open arXiv 2402.01680 and confirm that the section headings you plan to cite match the paper's own structure, since the README states the paper list and the paper are updated on different schedules. Second, check the repository directly for a licence file, because the metadata available here does not list one. If both checks come back as you expect, the list is a reasonable index. If you need working code today, pick one of the frameworks it links to and start there instead.

Editorial conclusion

Adopt this repository if you need a pre-sorted entry point into LLM multi-agent literature and you are willing to follow the arXiv links yourself. Do not adopt it if you need runnable code, benchmark harnesses or a licence you can rely on for redistribution: the repository supplies none of these, and the metadata retrieved here lists no licence at all. Before citing it, open arXiv 2402.01680 and confirm that the taxonomy in the paper matches the section headings you are citing, because the README states the paper list is updated on a two-week cadence and the two can drift apart.

Official sources

1. Issues
2. Project website
3. README
4. taichengguo/LLM_MultiAgents_Survey_Papers on GitHub