The cloud Agent is an extension of local capabilities. Handle things locally first, but when local is unsuitable or unable to complete the task, delegate to the cloud.
Capability boundaries between cloud and local
Cloud-only (local cannot do it)
- Isolated code sandbox: Safely execute arbitrary code (Python/Node/Shell) inside a Docker container without affecting the user's system
- Project build and publish: Build the project in the sandbox + package a Docker image + deploy to a public URL that users can access directly
- 24/7 continuous operation: Cloud scheduled tasks are not affected by whether the computer is powered on, and support three scheduling modes: cron/interval/once
- Cloud file storage: S3/OSS persistence, generating presigned download links
Local-only (cloud cannot do it)
- Desktop UI automation (peekaboo / pywinauto / Claude Computer Use)
- Local application control (open WeChat, operate Feishu, Apple Mail/Calendar, etc.)
- Screenshot + OCR, clipboard read/write
- Direct local file access (organize the desktop, move files)
Available on both ends (not a reason to delegate to the cloud)
- Web search: Locally there is DuckDuckGo/Brave/Tavily/arXiv/paper search, so daily searches can be done locally
- Web content extraction: Locally there is Jina Reader, Crawl4AI
- Knowledge base retrieval: Locally there is FTS5 + semantic search + RAG
- File processing: Locally you can directly read and write PDF/Excel/Word
When to delegate to the cloud
If anyone of the following reasonsis met, you can delegate:
Reason 1: Continuous operation — the task may take longer than this power-on session
Reason 2: Sandbox execution — you need an isolated environment to run code or build and publish a project
Reason 3: Deep research — long-duration, multi-round deep analysis that is not afraid of shutdowns
The advantage of cloud deep research is not "being able to search" (local can too), but rather:
- Not limited by power-on time, can run for hours
- Multi-round iterative search + analysis + search again, advancing autonomously
- Suitable for heavy outputs such as complete reports and comparative analyses
Scenarios where you must absolutely not delegate to the cloud
How to use
Basic call:
cloud_agent(task="调研 AI Agent 赛道最新动态并写一份分析报告")
With context:
cloud_agent(
task="基于以下竞品列表做深度调研",
context="竞品: Manus, Devin, Bolt..."
)
With files (PDF / Excel / images, etc.):
cloud_agent(
task="分析这份行业报告并提取关键数据",
files=[{"file_url": "/path/to/report.pdf", "file_name": "行业报告.pdf", "file_type": "application/pdf"}]
)
Local files are automatically uploaded to the cloud and then passed to the cloud Agent for processing.
Execution characteristics
- Streaming output: The cloud returns streaming by default, and progress is pushed in real time during execution (thinking -> tool call -> generating text)
- Execution time: Simple tasks 10-30 seconds, deep research 1-3 minutes, sandbox project build 2-5 minutes
- File support: Files such as PDF, Excel, CSV, and images can be passed in; local files are automatically uploaded to cloud S3
- Network dependency: Requires a network connection; when unreachable, the tool will return a clear error message
- Result format: Returns plain text results; if local files (PPT / Excel) need to be generated, this should be done locally