AgentsMeetRL plugin
The agents-meet-rl skill for Claude Code: a handbook on RL training of AI agents, failure diagnosis and framework choice.
- 1,880 ★ on the repository
- Last commit
Minimal access
Why this level
- The skill consists of reference documents and a project database. While working the agent may read your logs and code, but it runs nothing itself.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/agentsmeetrlDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill consists of reference documents and a project database. While working the agent may read your logs and code, but it runs nothing itself.
This is third-party code. Review the repository files before installing.
What it does
AgentsMeetRL began as a list of open-source repositories where AI agents are trained with reinforcement learning and now also ships as a plugin with one skill, `agents-meet-rl`. The skill turns the list into a handbook: the agent explains mechanisms, compares implementations, helps wire tool tasks and diagnoses causes of training and evaluation failures. Questions are routed by directory. `mechanisms/` holds write-ups on the trajectory contract, token and mask alignment, multi-turn objectives and async correction. `implementations/` has cards for slime, veRL and AReaL from source snapshots. The `problems/` directory groups symptoms by training, evaluation and research workflow. A machine-readable database covers 411 projects as of 10 October 2026. Each answer rests on specific papers and repositories. The skill separates derived reasoning, paper claims, source-inspected behavior and measured quality.
Who it is for. Researchers and engineers training LLM agents with reinforcement learning (GRPO, PPO, DAPO) and choosing between frameworks, environments and algorithms.
Good fit when
- The reward of a search agent being trained is flat while the evaluation metric drops.
- An RL framework must be chosen for a multi-turn tool-using agent.
- Entropy or response length changes sharply and the cause has to be found.
- Exported trajectories are checked offline for tool-call parse errors.
- An experiment is planned from scratch: baselines, ablations and compute budget.
Not a fit when
- The task is ordinary supervised fine-tuning: the skill excludes it.
- Classical RL without an LLM is needed: the skill excludes it.
- A live code or run check is wanted but the agent has no access to files and logs.
Example request
My GRPO agent's reward is flat while evaluation keeps dropping. Go through possible causes and say which logs to look at first.Limitations
The skill covers agentic LLM reinforcement learning and deliberately excludes ordinary supervised fine-tuning and classical RL without language models. The corpus was taken on 10 October 2026 and the engineering review was on 6 November 2026, so newer repositories are absent. The author warns the list was built by code analysis with AI agents and may contain inaccuracies, and states that source inspections do not prove training quality. Without access to code, logs and trajectories the agent gives a list of checks rather than a conclusion.
How to disable. Remove the plugin in the `/plugin` window or, for a manual install, the `~/.claude/skills/agents-meet-rl` directory, and start a new Claude Code session.
FAQ
How do I install the skill?
As a plugin: `/plugin marketplace add thinkwee/claude-plugins` and `/plugin install agents-meet-rl@thinkwee`. Manually: `git clone https://github.com/thinkwee/AgentsMeetRL` and `cp -r AgentsMeetRL/skills/agents-meet-rl ~/.claude/skills/`.
Do I have to call the skill with a command?
No. Once installed, Claude Code triggers the skill itself when a question concerns reinforcement learning of agents, for example about GRPO or reward.
Which problems does it cover?
Flat reward, KL and entropy spikes, tuning GRPO, PPO and DAPO, tokenization drift, tool-call parse errors, LLM-judge inconsistency and benchmark contamination.
What are implementation cards?
Versioned descriptions of slime, veRL and AReaL based on source snapshots. They help compare how each framework is built inside.
What is the corpus based on?
The README list of open-source repositories with real agent-training code and a machine-readable database of 411 projects. The list can be read without the skill.
Discussion
No questions or reviews yet.
Related
PluginsA skill for the "Early Experience" method (ICML 2026) that helps an agent prepare reward-free training data for language agents.
AI Builder Club Skills
AI Builder Club — Skills
A Claude Code plugin marketplace to make a repo agent-ready and run compounding agent loops
A plugin with one skill for designing, debugging and evaluating AI agents, based on Bojie Li book on agent engineering.
AI Research Skills
AI Research Skills Library
98 skills for ML research, from ideation and model training to inference, evaluation and paper writing
Anthropic skill for Claude Code that audits a multi-agent system and, by measurement, folds subagents into one loop with skills.
Awesome LLM apps
Awesome LLM Apps
A collection of over 100 AI agent and RAG app templates, including a section of skills for coding agents
README in short
The inclusion criterion is that a project has at least multi-turn interaction or tool use. Sections cover base frameworks, search and RAG, code and SWE, web and GUI, rewards and training, safety and environments.