TU2021

TU2021 / D2Skill-AgenticRL

Public

Dynamic dual-granularity skill bank for agentic RL, jointly evolving policy and skills to improve long-horizon decision making in agentic tasks.

arxiv.orgabs2603.28716
16
0
69% credibility
Found Apr 05, 2026 at 16 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Codebase implementing a dynamic skill bank for training reinforcement learning agents in environments like ALFWorld (household tasks) and WebShop (online shopping).

How It Works

1
🔍 Discover smart agent training

You find this project while reading about new ways to teach AI agents complex tasks like household chores or online shopping.

2
📥 Grab the code

Download the project files to your computer and prepare the basic tools it needs to run.

3
⚙️ Set up your workspace

Create a personal settings file to connect any helpful services your agent might use.

4
🏠 Add practice worlds

Bring in simulated homes and online stores where your agent can learn real-world skills.

5
🚀 Launch the skill helper

Start a background helper that recalls useful past actions to guide your agent.

6
🎯 Begin agent training

Kick off sessions where your agent practices tasks, improving with each try.

✅ Agent masters skills

Your AI now handles chores like cleaning or shopping confidently, ready for bigger challenges.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 16 to 16 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is D2Skill-AgenticRL?

D2Skill-AgenticRL is a Python framework for agentic RL that builds a dynamic dual-granularity skill bank, jointly evolving policies and skills to improve long-horizon decision making in agentic tasks like ALFWorld and WebShop. Developers get a setup for training agents with retrieval-based skills, handling environments via bash scripts that launch embedding services and run experiments. It tackles sparse rewards in extended tasks by dynamically retrieving and refining skills on-the-fly.

Why is it gaining traction?

Its dynamic workflow stands out by blending task-level and step-level skill retrieval, boosting performance on agenticrl benchmarks without manual skill engineering. Features like evolving skill banks and A/B baselines for ablation make experimentation straightforward, appealing to devs tweaking long-horizon agents. Early adopters note gains in success rates via jointly trained policies.

Who should use this?

RL engineers building agentic systems for household tasks or web navigation, like ALFWorld pick-and-place or WebShop shopping agents. Suited for researchers prototyping dynamic prompts in long-horizon setups, or teams needing a skill bank for decision-heavy sims.

Verdict

Worth forking for agentic RL prototypes if you're in long-horizon research—setup is solid with env scripts—but low maturity (16 stars, 0.699999988079071% credibility) means expect tweaks to docs and deps like vllm. Test on small batches first.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.