[Community Package] strands-sglang: SGLang model provider with Token-In/Token-Out (TITO) support for agentic RL training

[Lawhy]

Community Package

Hi team, I’m excited to share that we’ve released strands-sglang, a community package that integrates with SGLang’s native /generate endpoint to support token-in / token-out agentic RL training.

This package helps bridge a key gap in using agentic scaffolding for RL by enabling explicit token management, customizable tool-call parsing, and flexible agent loops—without tightly coupling agents to the underlying training infrastructure. With strands-sglang, users can implement training-ready agent loops in just a few lines of code, making it significantly easier to prototype and scale agentic RL workflows.

• PyPI: https://pypi.org/project/strands-sglang/
• GitHub: https://github.com/horizon-rl/strands-sglang

Related to Issue strands-agents/sdk-python#1368

Proposed Solution

strands-sglang

Use Case

strands-sglang targets agentic RL training with training frameworks like Slime and VeRL which utilize SGLang as the rollout engine

Alternatives Solutions

No response

Additional Context

No response

[Lawhy]

Potential Improvement: Currently, the token management is happening within the SGLangModel, making the model itself stateful. For a better design we would require the Agent class to take custom parameters, e.g., the token manager, for better modularity.

[Lawhy]

Hi @zastrowm, may I ask if this package suits the community package's standard?

[zastrowm]

@Lawhy, yes.

We don't have the process fully documented yet, but if you create a pull-request to this project:

• Create a new file under the community/model-providers directory;
• Follow existing conventions (see this as an example):
  • The general layout is:
    • Quick Intro
    • Installation (using pip)
    • Usage (with code sample)
    • Configuration (maybe with Client or Model configuration)
    • Troubleshooting or advanced usage [if needed]
    • References (links)
• Update this file to reference this new file

[Lawhy]

Hi @zastrowm, thanks for your guidance, please see the PR #419!

[dk67604]

Hi @Lawhy Following your example I created community package for vllm: https://github.com/agents-community/strands-vllm. Thanks for putting a very good example.

[Lawhy]

Hi @dk67604, thanks for putting VLLM support, we will mention it in Strands-SGLang as well! One thing I noticed that is you seem to rely on VLLM's own tool parsing logic. This can be fragile for training as we want strict tool parsing during training.

[dk67604]

Thanks @Lawhy I appreciate, it was great collaborating with you on this.