• 運算
  • 客戶
  • 價格
登入
More Blog Posts
XDiscordLinkedInYouTube

產品

  • GPU
  • MaaS
  • Studio

開發者

  • 模型總覽
  • 技術文件
  • 詞彙表

公司

  • 關於我們
  • 部落格
  • 活動
  • 合作夥伴
  • 新創計劃
  • 職涯
  • 大使計畫
  • 使命與願景

熱門模型

    掌握 AI 最新動態

    提交即表示您瞭解我們會收集並使用您提交的資訊,其中可能包含個人資訊。

    XDiscordLinkedInYouTube

    Copyright ©2026 All rights reserved.

    隱私政策使用條款法律文件
    More Blog Posts
    Product

    Agentbox v2: built for agents that stay

    2026年10月22日

    Early access for Agentbox v2 is here. Agents can keep the same project files across multiple runs and pause compute while they wait. Storage for paused sandboxes is free during early access.

    An agent's work doesn't always end with its response. A user returns with feedback, a report needs another revision, or a coding task continues the next morning. The next run needs the files from the last one, but keeping the environment running through those gaps costs compute.

    Agents can still create a sandbox for a single task, collect the result, and delete it. V2 adds the option to keep a project workspace between runs.

    The agent pauses when it needs to wait

    When an agent is done for now, it can save its progress and pause the sandbox. Compute stops, and saved files stay available.

    When the next request arrives, your application resumes the sandbox so the agent can continue working from those files. The sandbox keeps the same ID, files, and configuration.

    Resume restarts the environment and reruns its startup command. It does not restore memory or continue running processes. Progress needed by the next run must be saved to files before pausing.

    The agent controls its workspace through code

    Create, execute, upload, download, pause, resume, and delete are available through the SDK. Agents can use these operations as tools while they work.

    sandbox.execute("pytest")       # agent runs its task
    sandbox.pause()                 # done for now — compute stops, files stay
    
    # ...hours later, same sandbox, same files
    sandbox.resume()
    sandbox.execute("pytest")       # picks up where it left off

    If the agent runs inside the sandbox it pauses, an external application must initiate the resume. V2 also provides a maximum runtime set at launch, optional idle auto-pause, and deletion from any sandbox state.

    Early access

    Use a fresh sandbox for a one-off task, or keep the project files for the next run. Agentbox v2 supports both.

    Compute is usage-based. Paused sandboxes consume no compute, and paused-sandbox storage is free during early access.

    → Request early access


    Build AI Without Limits

    GMI Cloud helps you architect, deploy, optimize, and scale your AI strategies

    Ready to build?

    Explore powerful AI models and launch your project in just a few clicks.

    Get Started