AgentsMeetRL: a curated index of open source RL training code for LLM agents

The gap AgentsMeetRL is trying to fill

Agentic reinforcement learning sits at the intersection of two fast-moving literatures, and the repositories are scattered. A paper may describe a multi-turn tool-use training loop, but the code lives in a lab repo with no announcement, or appears months after publication. AgentsMeetRL is an attempt to keep a single index of those repositories. The README states the project summarizes open source repositories for training LLM agents using reinforcement learning, and it gives an explicit inclusion rule: a project qualifies as an agent project if it has at least one of multi-turn interactions or tool use. That second clause is why tool-integrated reasoning work is counted here rather than in a general reasoning list. The audience is narrow and identifiable: researchers and engineers choosing an RL framework, a reward scheme, or a training environment for an agent, who want to see what other projects actually did rather than what a paper abstract claims.

Sixteen categories, and what each one is for

The taxonomy is the substance of the list. Base Framework covers general-purpose RL training frameworks for LLM agents, and the README names veRL, OpenRLHF and trl as examples. General/MultiTask covers agent systems trained or evaluated across multiple tasks. Search & RAG covers search-augmented reasoning agents that call retrieval tools. Web & GUI covers agents that drive browsers, mobile or desktop GUIs, or operating systems. Tool-Use covers agents trained to invoke external tools such as APIs, code executors or MCP. Code & SWE covers software engineering and code generation agents. Reasoning covers tool-integrated or multi-turn reasoning for math, QA and visual tasks. Multi-Agent RL covers collaboration, negotiation and credit assignment. Memory, Embodied, Domain-Specific, Reward & Training, Safety, VLM Agent and Self-Evolution fill out the rest, with Environment collecting benchmarks, gyms and sandboxes. Two of these categories carry honest caveats in the README: Self-Evolution is marked as a definition still evolving in the community, and the Environment category is the largest by the badge counts, which suggests environments are where the current bottleneck and the current churn both sit.

How entries are produced, and why that matters

The README is unusually direct about method. The list is based on code analysis from open source repositories using LLM coding agents, and the authors state it may contain unfaithful cases, adding that although manually reviewed there may still be omissions. That is a meaningful disclosure. It means each entry is a reading of a repository rather than a reproduction of its results, and the review step is described as manual but not exhaustive. The update notes reinforce the point: the 2026-08 entry says every repo was opened and confirmed to contain real RL-training or executable-environment code, and lists papers left out because their code is unreleased, including Qwen-UI-Agent, Qwen-CUA, UI-Mate, SearchMaster and others. The same note observes that no qualifying new Safety, Embodied or Multi-Agent RL repositories appeared in that window, describing that crop as uniformly SFT-only, inference-only, or code-withheld. Treat the list as a map of what code exists, not as a verification that the code trains anything to a useful standard.

Reward types and the per-table technical detail

The list enumerates reward types as a controlled vocabulary: External Verifier (a compiler or math solver), Rule-Based (for example a LaTeX parser with exact match scoring), Model-Based (a trained verifier or reward LLM), and Custom. This is the part most likely to save time, because reward design is where agentic RL projects diverge most and where a paper's description is often least concrete. The README also says the project focuses on the reinforcement learning frameworks, RL algorithms, rewards and environments that projects depend on, and that technical details are reachable through a Click to view technical details control under each table. The categories also carry counts in the badge row at the top of the README: Environment at 64, Search & RAG at 50, Web & GUI at 32, VLM Agent at 30, Base Framework at 29, Tool-Use and Code & SWE at 26 each, General at 21, Reasoning and Self-Evolution at 18 each, Multi-Agent at 14, Domain-Specific at 12, Reward & Training at 11, Safety at 9, Memory at 8, Embodied at 7. Those are inventory counts, not quality signals, and they date from the last update.

Getting to the list and contributing to it

There is nothing to install. The primary artifact is the repository README plus a static site. The README links an Interactive Dashboard at https://thinkwee.top/amr/, which is also the repository homepage, and the site is the intended way to browse the tables. The repository's primary language is listed as HTML, which is consistent with a generated site rather than a Python package; no releases have been published, so there is no versioned artifact to pin. Contribution is by issue or pull request: the README asks readers to report errors through issues or PRs and says submissions of new projects are welcome at any time. Citation is handled through the Cite this repository button in the right sidebar rather than a BibTeX block in the README. If you want to check whether a specific project is covered, the fastest route is the dashboard's category tables rather than a text search of the README, since the README shows the taxonomy and the update notes but not the full per-project tables.

Where the list is the wrong tool

Three failure modes are worth naming. First, the list is not a benchmark. It records which framework, algorithm, reward type and environment a project uses; it does not report training curves, success rates or compute budgets, and the README makes no such claim. If you need to know whether a method beats a baseline on a specific task, this repository will not answer that. Second, the inclusion rule excludes a large body of work by design. A single-turn RLHF-style project with no tool use and no multi-turn interaction is out of scope even if it is a strong agent backbone, and papers whose code is unreleased are held back, which the update notes show is a recurring and deliberate decision. Third, the analysis layer is a model reading code. A repository can be present and still be described inaccurately in its reward or algorithm field; the README's own disclaimer says so. Anyone citing a per-project detail from this list should confirm it against the linked source.

What you would use instead, and how it differs

The obvious alternative is a general awesome list for LLM agents or for reinforcement learning. The difference is the filter, not the format. A general agent list indexes by application domain and accepts inference-only, prompt-engineering and SFT projects alongside training code; AgentsMeetRL requires multi-turn interaction or tool use and, per the update notes, requires that the repository contain real RL-training or executable-environment code before it is listed. A general RL list indexes by algorithm family and is indifferent to whether the policy is an LLM agent. The cost of AgentsMeetRL's tighter filter is coverage: it will not help you find an agent framework that is excellent but only supports supervised fine-tuning, and it will not help you find a paper worth reading whose code is still private. If your question is which reward function or environment to reuse for an agent training run, the narrower list is the better starting point. If your question is what the field looks like overall, it is not.

Licence, maintenance and what to verify before adopting

No licence is stated in the repository metadata supplied here, and the README does not name one. That is a real gap if you intend to reuse the taxonomy, the generated site, or any table content in your own work: without a licence file, the default position is that no rights are granted, and this article cannot give legal advice on what that means for your situation. Maintenance looks active rather than dormant. The most recent push is dated 2026-08-28, the README states a last-updated date of 2026-08-26, and the update log shows monthly batches through 2026-06, 2026-07 and 2026-08, with the June batch alone adding 43 repositories. The cost of following it is therefore low but not zero: the categories are re-counted and the tables rewritten each cycle, so a link you saved may move between categories or be superseded. Before you adopt anything from the list, verify two things at the source repository: that the training code is present and runnable, and which licence that repository carries, because AgentsMeetRL does not carry that information for you.

Editorial conclusion

Use AgentsMeetRL if you need to find open source RL training code for agent behaviours such as multi-turn interaction or tool use, and you want the reward type and environment for each entry in one place. Do not use it as a source of benchmark numbers, as a survey of unreleased work (the README deliberately excludes papers whose code is not out), or as a dependency in a build. Before relying on any entry, open the linked repository and confirm two things yourself: that the code is actually present and trainable, and what licence it carries, since AgentsMeetRL does not state one for itself and the README's own disclaimer says the code analysis may contain unfaithful cases.

Official sources

1. Issues
2. Project website
3. README
4. thinkwee/AgentsMeetRL on GitHub