A curated map of 312 projects training LLM agents with RL

What it does

AgentsMeetRL catalogs open-source repositories that train LLM agents via reinforcement learning. The maintainer uses LLM coding agents to scan codebases, then manually reviews the results. Projects are sorted into 16 categories — from base frameworks like veRL and OpenRLHF to niche buckets such as Embodied, Safety, and Self-Evolution.

The interesting bit

The list is packaged as a Claude Code skill (agents-meet-rl) that auto-invokes when you complain about training pathologies — flat GRPO rewards, KL blow-ups, tool-call parse failures — and grounds answers in specific papers and repos from the corpus. It is essentially a machine-readable literature review wired into your IDE.

Key highlights

• 312 projects as of May 2026, with taxonomy tracking frameworks, reward types, and environments per entry
• 16 categories including VLM Agent, Multi-Agent RL, and Reward & Training
• Interactive dashboard at thinkwee.top/amr/ for browsing the corpus
• Monthly updates; recent batches added 67 repos (Apr 2026) and 17 more (May 2026)
• Reward-type enumeration covers External Verifier, Rule-Based, Model-Based, and Custom

Caveats

• The curation pipeline relies on LLM coding agents, so the README warns of possible “unfaithful cases” and omissions despite manual review
• Some entries lack paper links or released code (e.g., CoEvolve is flagged “Under Review”)
• The “Self-Evolution” category carries a disclaimer that the definition is still evolving in the community

Verdict

Worth bookmarking if you are building or benchmarking agentic RL systems and need to compare framework choices. Less useful if you want runnable code or tutorials — this is a reference index, not a starter kit.