If you’re building AI agents, RLHF pipelines, or training infra: give trl-training a spin in v0.29.0 and let us know what you think. We are working at improving it! 🤗
The era of agent-native ML tooling is here 🚀
🔗 github.com/huggingface/...
Under the hood:
- A dedicated trl-training skill definition
- CLI-first design
- Clean integration with existing TRL commands
- Structured documentation so LLM agents can reason over it
What does this mean?
Instead of “figuring out” how to train with TRL…
Agents now have a clearly defined interface to:
✅ Run SFT
✅ Run DPO
✅ Run GRPO
✅ Launch training jobs
✅ Work with the TRL CLI safely
All through a structured SKILL.md.
🚀 Big news for TRL users with AI agents:
We just released the trl-training SKILL in TRL v0.29.0 🤗
Agents can now use the TRL CLI as a structured capability:
5 years already working in democratizing AI 🤗
Grateful to be part of such an awesome team making it happen every day.
Latest smolagents release supports GPT-5: build agents that think, plan, and act.
⚡ Upgrade now and put GPT-5 to work!
🚀 smolagents v1.21.0 is here!
Now with improved safety in the local Python executor: dunder calls are blocked!
⚠️ Still, not fully isolated: for untrusted code, use a remote executor instead: Docker, E2B, Wasm.
✨ Many bug fixes: more reliable code.
👉 github.com/huggingface/...
🔥 This feature is live in smolagents v1.20.0!
Try it out.
Break things. Extend it. Give us feedback.
Let's build safer, smarter agents; together 🧠⚙️
👉 github.com/huggingface/...
#smolagents #opensource #AIagents #WebAssembly #Deno #Pyodide #HuggingFace #AgenticAI
Check out the PR 👉 github.com/huggingface/...
Let's reimagine what agent-driven code execution can look like: remote, safe, and Wasm-powered.
👷 Want to help?
This is a foundation we're inviting the community to build on:
• Add feature support
• Expand compatibility to new use cases
• Help make agents even more secure + capable
Contribute, test, break things, and PR back 🚀
🙅♂️ Why no network calls (for now)?
Because Pyodide in Node.js needs experimental flags for stack switching, and some APIs (like urllib3) aren’t yet compatible out of the box.
But this is just the beginning.
🚧 What’s supported right now?
✅ Basic Python logic
✅ Standard control flow
✅ Limited stdlib support
❌ No HTTP requests (yet)
❌ No advanced I/O or external packages
It's early: a solid MVP for secure use cases. 💡
🔐 Why WebAssembly?
WebAssembly = portable, fast, and secure.
It enables:
• Safe remote execution of Python code
• Zero access to the host machine
• Consistent, deterministic environments
• Serverless + edge-compatible agent workflows
💥 Big milestone just dropped in smolagents:
The CodeAgent can now execute Python code remotely in WebAssembly, powered by Pyodide + Deno!
Your agent-generated Python runs in a secure, sandboxed Wasm runtime: no local execution required.
Here's why this changes the game 🧵👇
7/
🧵 That’s a wrap for smolagents v1.19.0!
Grab the latest, test the improvements, and tell us what you build.
👇👇👇
github.com/huggingface/...
#smolagents #AIagents #OSS #HuggingFace
6/
📚 Docs Game: Leveled Up
* ✨ Quickstart got tool usage examples
* 💡 Added open in Colab buttons!
* 🧠 Reference docs now include AgentMemory, better docstrings
5/
🐛 Bug Fixes Blitz
From final answer logic to Docker logs, this release patched up:
* Planning logic bugs 🧠
* Type-safe wrappers 🛡️
* Executor output matching 🧵
* Prompt fixes for additional_args 📝
4/
📋 Agent Output Tracking
* CodeAgent outputs are now stored in ActionStep
📂 More structured traceability of actions & results!
3/
📡 Streaming Just Got Smarter
* Streaming event aggregation was moved off the Model class
➕ Cleaner structure, better separation of concerns
2/
🎛️ UI Improvements
* New reset_agent_memory feature in GradioUI 💾
- Handy for debugging or starting fresh mid-run!
1/
🧠 Agent Upgrades
* You can now use managed agents with ToolCallingAgent 🔧
* Agent cleanup supports context managers: easier resource handling!
* Switched code output tags to XML for clarity
🚀New Release: smolagents v1.19.0 is out!
This one's packed with agent upgrades, smoother streaming, better docs, and bug squashing galore. 🧵👇
New in smolagents v1.17.0:
- Structured generation in CodeAgent 🧱
- Streamable HTTP MCP support 🌐
- Agent.run() returns rich RunResult 📦
Smarter agents, smoother workflows.
Try it now: github.com/huggingface/...
🚀 Experiment with Tiny Agents: minimalist LLM agents powered by async MCP tools, inspired by @julien_c
🔍 A side-by-side comparison of tool-calling vs. code agent
💡 See how code agents handle complex workflows in one step, reducing model calls and boosting efficiency!
New in smolagents v1.16.0:
🔍 Bing support in WebSearchTool
🐍 Custom functions & executor_kwargs in LocalPythonExecutor
🔧 Streaming GradioUI fixes
🌐 Local web agents via api_base & api_key
📚 Better docs
👉 github.com/huggingface/smolagents
Code Agents generate full code blocks and execute them in one go, instead of calling tools one-by-one. Learn how to build smarter, faster agents in a new @deeplearningai.bsky.social course by @hf.co.
Build with smolagents → sandboxed, reliable, powerful.
🔗 www.deeplearning.ai/short-course...
Full release 👉 github.com/huggingface/...
#smolagents #LLM #AgenticAI
smolagents v1.14.0 is out! 🚀
🔌 MCPClient: A sleek new client for connecting to remote MCP servers, making integrations more flexible and scalable.
🪨 Amazon Bedrock: Native support for Bedrock-hosted models.
SmolAgents is now more powerful, flexible, and enterprise-ready. 💼
International Women's Day (IWD) 2025: www.internationalwomensday.com