[Roadmap] Uni-Agent 26Q3 Roadmap

[yyDing1]

Refactor

First, uni-agent will focus on refactoring at the framework level to make the framework more user-friendly.

Then, building upon the new feat, we plan to implement the following features.

Performance optimization

These performance optimization features for Agentic RL will be initially incubated and validated within the uni-agent repository for rapid iteration. Once the architecture matures and stabilizes, we plan to upstream/migrate them into the main verl repository.

• Communication Bottleneck Optimization for Multi-Modal Agentic RL
• Hybrid Sequence Parallelism for Long-Context Agentic RL

Gateway

• Multi-modal
• Support Anthropic API
• Support sub-agent
• Support rewrite assistant message
• Support prefix-tree trajactory storage ([RFC] Prefix Trie for Multi-trajectories Storage for AgentGateway #51)

Agent

• Claude Code Harness
• Support multi-agent/multi-policy training
• Claw-style agent, e.g., openclaw
• AI4S agents, e.g., claude science

Task

• Support GUI Agent (e.g., OS-World)

Sandbox

RL recipe

[bbbearxyz]

Hi @yyDing1, thanks for putting together this roadmap!

I'm particularly interested in the Agent section and would love to contribute to it. Before diving in, I'd like to check on the current status:

1. Are the Agent items (Claude Code Harness, multi-agent/multi-policy training, Claw-style agent, AI4S agents) all still unimplemented, or is any of them already in progress? I'd like to avoid duplicating ongoing work.
2. Among these, which one would you consider the highest priority or the best entry point for a new contributor?
3. Is there an existing design doc or RFC for any of these items? If not, would you prefer that I start by drafting a proposal in a sub-issue before writing code?
4. What's the recommended way to claim an item — commenting here, or opening a dedicated tracking issue? And is there a communication channel (Slack / Discord / WeChat group) where design discussions happen?

Happy to start with whichever item is most useful to the project. Looking forward to your guidance, thanks!

[yyDing1]

Hi @bbbearxyz, thanks for your interest and willingness to contribute!

We are currently landing a substantial refactor of Uni-Agent in #83 and the new docs. With the new architecture, integrating additional agent harnesses should be relatively straightforward, and contributions adding new agents are very welcome.

For a first contribution, I would recommend integrating a concrete single-agent harness like a Claw-style agent. You can use the Claude Code implementation as a reference. You can also add something you’d like to build to the roadmap. PR #83 already includes Claude Code as a black-box agent demo. Another black-box harness, mini-swe-agent, is also being worked on in #73.

For discussion, GitHub Issues are preferred for design decisions so the discussion remains public and searchable, while the verl-project community channels are also available for real-time discussion.

Thanks again, and look forward to your contribution!