|
| 1 | +--- |
| 2 | +title: "Building the Open Agent Ecosystem Together: Introducing OpenEnv" |
| 3 | +thumbnail: /blog/assets/openenv/thumbnail.png |
| 4 | +authors: |
| 5 | +- user: spisakjo |
| 6 | + guest: true |
| 7 | + org: openenv |
| 8 | +- user: darktex |
| 9 | + guest: true |
| 10 | + org: openenv |
| 11 | +- user: zkwentz |
| 12 | + guest: true |
| 13 | + org: openenv |
| 14 | +- user: mortimerp9 |
| 15 | + guest: true |
| 16 | + org: openenv |
| 17 | +- user: Sanyam |
| 18 | + guest: true |
| 19 | + org: openenv |
| 20 | +- user: Hamid-Nazeri |
| 21 | + guest: true |
| 22 | + org: openenv |
| 23 | +- user: Pankit01 |
| 24 | + guest: true |
| 25 | + org: openenv |
| 26 | +- user: emre0 |
| 27 | + guest: true |
| 28 | + org: openenv |
| 29 | +- user: lewtun |
| 30 | +- user: reach-vb |
| 31 | +--- |
| 32 | + |
| 33 | +# Building the Open Agent Ecosystem Together: Introducing OpenEnv |
| 34 | + |
| 35 | +With tools like [TRL](https://github.com/huggingface/trl), [TorchForge](https://github.com/meta-pytorch/torchforge) and [verl](https://github.com/volcengine/verl), the open-source community has shown how to scale AI across complex compute infrastructure. But compute is only one side of the coin. The other side is the developer community; the people and tools that make agentic systems possible. That’s why Meta and Hugging Face are partnering to launch the [OpenEnv Hub](https://huggingface.co/openenv): a shared and open community hub for agentic environments. |
| 36 | + |
| 37 | +Agentic environments define everything an agent needs to perform a task: the tools, APIs, credentials, execution context, and nothing else. They bring clarity, safety, and sandboxed control to agent behavior. |
| 38 | + |
| 39 | +These environments can be used for both training and deployment, and serve as the foundation for scalable agentic development. |
| 40 | + |
| 41 | +## The Problem |
| 42 | +Modern AI agents can act autonomously across thousands of tasks. However, a large language model isn’t enough to get those tasks to actually run — it needs access to the right tools. Exposing millions of tools directly to a model isn’t reasonable (or safe). Instead, we need **agentic environments**: secure, semantically clear sandboxes that define exactly what’s required for a task, and nothing more. These environments handle the critical details: |
| 43 | +- Clear semantics about what a task needs |
| 44 | +- Sandboxed execution and safety guarantees |
| 45 | +- Seamless access to authenticated tools and APIs |
| 46 | + |
| 47 | +## The Solution |
| 48 | +To supercharge this next wave of agentic development, Meta-PyTorch and Hugging Face are partnering to launch a [Hub for Environments](https://huggingface.co/openenv): a shared space where developers can build, share, and explore OpenEnv-compatible environments for both training and deployment. The figure below shows how **OpenEnv** fits in the new post-training stack being developed by **Meta**, with integrations for other libraries like **TRL**, **SkyRL**, and **Unsloth** underway: |
| 49 | + |
| 50 | +<p align="center"> |
| 51 | + <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/openenv/rl-stack.png" alt="rl_stack" width="750"/> |
| 52 | +</p> |
| 53 | + |
| 54 | +Starting next week, developers can: |
| 55 | +- Visit the new [Environment Hub](https://huggingface.co/openenv) on Hugging Face where we will seed some initial environments |
| 56 | +- Interact with environments directly as a Human Agent |
| 57 | +- Enlist a model to solve tasks within the environment |
| 58 | +- Inspect which tools the environment exposes and how it defines its observations |
| 59 | +- Every environment uploaded to the Hub that conforms to the OpenEnv specification automatically gains this functionality — making it fast and easy to validate and iterate before running full RL training. |
| 60 | + |
| 61 | +Alongside this, we’re releasing the [OpenEnv 0.1 Spec (RFC)](https://github.com/meta-pytorch/OpenEnv/blob/main/rfcs/002-env-spec.md) to gather community feedback and help shape the standard. |
| 62 | + |
| 63 | +## The RFCs |
| 64 | +In the current state of the repository, environment creators can create environments using `step()`, `reset()`, `close()` APIs (part of RFCs below). A few examples on how to create such environments can be seen [here](https://github.com/meta-pytorch/OpenEnv/tree/main/src/envs). Environment users can play with local Docker based environments for all environments already available in the repo. Following RFCs are under review: |
| 65 | + |
| 66 | +- [RFC 001:](https://github.com/meta-pytorch/OpenEnv/blob/main/rfcs/001-abstractions.md) Establish architecture for how the core components like Environment, Agent, Task, etc. are related |
| 67 | +- [RFC 002:](https://github.com/meta-pytorch/OpenEnv/blob/main/rfcs/002-env-spec.md) Propose basic env interface, packaging, isolation and communication w/ environment. |
| 68 | +- [RFC 003:](https://github.com/meta-pytorch/OpenEnv/blob/main/rfcs/003-mcp-support.md) Propose encapsulation of MCP tools through environment abstraction and isolation boundaries |
| 69 | +- [RFC 004:](https://github.com/meta-pytorch/OpenEnv/blob/main/rfcs/004-actions-as-tool-calls.md) Extend tool support to cover unified action schema covering tool calling agents as well as CodeAct paradigm. |
| 70 | + |
| 71 | +## Use cases |
| 72 | +- RL Post training: pull in environments across collections and use them to train RL agents with TRL, TorchForge+Monarch, VeRL etc. |
| 73 | +- Environment creation: build an environment and ensure that it interops with popular RL tools in the ecosystem, share with collaborators, etc. |
| 74 | +- Reproduction of SOTA methods: easily replicate methods like those from FAIR's [Code World Model](https://huggingface.co/papers/2510.02387) by integrating environments for agentic coding and software engineering. |
| 75 | +- Deployment: users can create an environment, train on the same environment and then use the same for inference too (the full pipeline) |
| 76 | + |
| 77 | +## What’s Next |
| 78 | +This is just the beginning. We’re integrating the OpenEnv Hub with Meta’s new **TorchForge RL library**, and collaborating with other open-source RL projects such as **verl**, **TRL**, and **SkyRL** to expand compatibility. |
| 79 | +Join us at the PyTorch Conference on Oct 23 for a live demo and walkthrough of the spec, and stay tuned for our upcoming community meetup on environments, RL post-training, and agentic development. |
| 80 | + |
| 81 | +👉 Explore the [OpenEnv Hub](https://huggingface.co/openenv) on Hugging Face and start building the environments that will power the next generation of agents. |
| 82 | + |
| 83 | +👉 Check out the 0.1 spec which can be found implemented in the [OpenEnv project](https://github.com/meta-pytorch/OpenEnv) → we welcome ideas and contributions to making it better! |
| 84 | + |
| 85 | +👉 Engage on [Discord](https://discord.gg/YsTYBh6PD9) and talk with the community about RL, environments and agentic development |
| 86 | + |
| 87 | +👉 Try it out yourself - We created a comprehensive [notebook](https://colab.research.google.com/github/meta-pytorch/OpenEnv/blob/main/examples/OpenEnv_Tutorial.ipynb) that walks you through an end to end example and of course you can easily pip install the package via [PyPI](https://pypi.org/project/openenv-core/). This notebook walks you through the abstractions we’ve built, along with an example of how to use existing integrations and how to add yours - Try it out in Google Colab! |
| 88 | + |
| 89 | +👉 Check out supporting platforms - [Unsloth](https://github.com/unslothai/unsloth), [TRL](https://huggingface.co/docs/trl/main/en/openenv), [Lightning.AI](http://Lightning.AI) |
| 90 | + |
| 91 | +Let's build the future of open agents together, one environment at a time 🔥! |
0 commit comments