AliyunConsoleAgent: Training Web Agents in Real-World Cloud Environments via Distillation and Reinforcement Learning

Researchers introduce AliyunConsoleAgent, a framework that trains cost-efficient web agents to automate documentation verification in cloud consoles through a combination of supervised learning from proprietary model trajectories and reinforcement learning in real cloud environments. The 32B parameter model achieves 63.52% success rate on a challenging benchmark, approaching proprietary frontier models at 92% lower inference cost.

AliyunConsoleAgent addresses a critical operational challenge for cloud platforms: the exponential labor cost of maintaining documentation accuracy as cloud consoles evolve. With an estimated 4 million annual inspections needed yet less than 1% manual coverage achieved, the problem represents substantial waste and risk for enterprises relying on outdated procedures. The framework demonstrates a pragmatic engineering approach to narrowing the capability gap between expensive proprietary models and open, deployable alternatives.

The technical innovation combines knowledge distillation from frontier models with reinforcement learning in deterministic, audited cloud environments. By using Terraform-based resource provisioning and rule-based reward models grounded in backend audit logs, the team solved a fundamental challenge in agent training: isolating the actual outcome signal from environmental noise. This methodological contribution has applications beyond cloud documentation verification, particularly for any task requiring agents to operate in complex, real-world systems where ground truth is verifiable but expensive.

For the AI industry, this work validates that smaller, open models can achieve near-frontier performance through careful training methodology rather than pure scale. The 92% cost reduction at near-parity performance suggests economic pressure on proprietary model pricing. For cloud providers like Alibaba, automation of documentation verification improves operational efficiency and customer experience. The framework's reliance on distillation from frontier models raises questions about long-term sustainability—as open models improve independently, dependence on proprietary knowledge diminishes. Enterprises should monitor whether similar cost-efficient agent frameworks emerge for other enterprise automation tasks, potentially reshaping the competitive landscape around AI service pricing.

• →AliyunConsoleAgent achieves 63.52% success rate, nearly matching 65.34% frontier model performance at 92% lower inference cost
• →Two-stage training paradigm combines supervised fine-tuning on distilled trajectories with reinforcement learning in deterministic cloud environments
• →Rule-based reward evaluation using backend audit logs prevents reward hacking and provides objective outcome judgment
• →Framework reduces documentation verification gap from 4 million annual inspections needed to <1% manual coverage
• →Success demonstrates smaller open models can match proprietary performance through distillation and RL optimization