AgentsSeptember 6, 2026via MarkTechPost
UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents
Why it matters
CUA-Lite removes infrastructure friction that has split the agent-training ecosystem into incompatible silos. Practitioners building or fine-tuning computer-use agents now have a standard platform for sandbox environments, training data, evaluation, and reinforcement learning — lowering the barrier to agent development and enabling broader experimentation.
Key signals
- CUA-Lite unifies agents, environments, traces, and evaluation framework behind one action space and data schema
- Replaces per-task virtual machines with Docker containers: 0.9 GB vs 4.1 GB (78% reduction)
- Addresses fragmentation: today's sandboxes, data formats, and RL frameworks are incompatible
- UC Berkeley led team
- Open-source platform
- Replaces OSWorld's infrastructure approach
The hook
UC Berkeley just unified the fragmented tooling for building computer-use agents — one action space, one data schema, Docker instead of 4GB VMs.
Training and benchmarking a computer-use agent needs four things — agents, environments, traces, and a framework to evaluate and train them — and all four ship in incompatible formats today. CUA-Lite, from a UC Berkeley led team, puts them behind one action space and one data schema, and replaces OS…