~/AI AGENTS/anthropic-shifts-claude-cowork-tool-execution-to-cloud-hosted-sandboxes

Anthropic Shifts Claude Cowork Tool Execution to Cloud-Hosted Sandboxes

Anthropic engineer Felix Rieseberg revealed that Claude Cowork updated its architecture to execute model inference and virtual machines inside isolated cloud sandboxes rather than on local user hardware. This change enables tasks to continue running in the background when a laptop is closed, preserves battery life, and allows users to trigger tasks from mobile devices. This evolution highlights key technical trade-offs when building autonomous AI agents that need to interact safely with user environments. Shifting execution to cloud sandboxes solves practical usability pain points like battery drain and session interruptions while preserving strict security boundary isolation. Each Cowork session now runs in its own isolated cloud sandbox without sharing state across sessions. When an agent requires access to a local resource like a file on the user's desktop, the desktop application handles that specific tool call to bridge the local and cloud environments.

## BACKGROUND

AI agents frequently execute code and perform multi-step workflows across files and web tools, requiring secure execution environments to prevent unauthorized system access. Anthropic originally bundled a lightweight virtual machine with the desktop app so tool execution happened locally on the user's hardware.

## REFERENCES

## KEYWORDS

#AI Agents#Systems Architecture#Cloud Computing#Security Sandboxing#Anthropic

$ subscribe --daily

Anthropic Shifts Claude Cowork Tool Execution to Cloud-Hosted Sandboxes | Daily News