# Dropbox Integrates AI Agents with Secure File System Controls Engineers developed an isolated workspace allowing AI agents to perform complex file tasks safely. - **Source:** [Dropbox Tech](https://dropbox.tech) - **Published:** 2026-11-05 - **Topic:** [technology](https://brifing.si/c/technology) - **Original article:** - **On Brifing:** https://brifing.si/s/f3791de9-7c11-4613-b396-cde0ff16aa59 ## Evolution of AI Agents Dropbox is transforming AI from a simple text-generation tool into an active agent capable of modifying files. Engineers Daniel Reiter Horn and Rich Chan developed this system by merging two internal experiments. The goal is to allow users to request complex tasks like folder organization or file renaming directly within their workspace. ## Two Distinct Experimental Approaches Daniel’s prototype focused on an isolated sandbox environment to run specialized 3D modeling software without local installation. Rich’s approach prioritized a graphical interface within the Dropbox desktop app to manage expense reports. Combining these projects created a unified system that balances powerful software execution with user-friendly permission controls and file access. ## Secure Isolated Computing Environments The system utilizes Riviera, a content-processing platform, to launch fresh virtual machines for every conversation. These environments run on Ubuntu and are isolated from the general internet to ensure security. By using QEMU snapshots, the system can spin up these workspaces in tens of milliseconds, providing a clean slate for each task. ## Strict Permission and Access To prevent unauthorized changes, all sensitive credentials remain outside the virtual machine. A proxy layer intercepts requests, verifying that the user has granted specific permissions before any file operation proceeds. This architecture ensures the agent cannot grant itself broader access, keeping the user in total control of their data and system. ## Structured Communication Protocols The interface communicates with the agent using the Codex App Server API, which exchanges structured JSONL messages. These messages are organized into threads, turns, and items to track every action. This structure allows the interface to display progress updates as cards, showing users exactly what the agent is planning and doing. ## User-Centric Planning and Approval For complex requests, the agent must propose a plan before taking action. The interface presents these options as cards, allowing users to choose specific parameters like organizing photos by date or location. The agent is blocked from modifying files until the user explicitly approves the proposed plan and grants the necessary permissions. ## Optimizing System Startup Speed Connecting to Dropbox files previously caused latency in workspace preparation. To improve performance, the system begins preparing the environment as soon as a user starts typing a request. By moving tools like rclone to a separate virtual disk, the team avoids rebuilding snapshots, ensuring the workspace is ready within 300 milliseconds. ## Precise Undo and Rollback Dropbox records which conversation generated specific file changes to enable precise rollbacks. If an agent makes an error, users can undo changes related to a specific turn without affecting other files. If a file was edited by another user, the system saves the restored version as a separate copy to prevent data loss. ## Cross-Harness Compatibility Adapters The team implemented an adapter layer to translate messages between different coding agent harnesses, such as Pi, and the Codex protocol. This design allows Dropbox to swap the underlying AI model or harness without rebuilding the entire user interface. Permissions remain enforced in the trusted host layer, ensuring consistent security across all models. ## Visibility into Agent Activity Users can monitor the agent’s progress through a checklist interface that displays active and completed steps. This transparency allows users to interrupt or steer the conversation at any time. By grouping activities under relevant plan steps, the system makes complex automated processes understandable without requiring the user to possess any coding knowledge. ## Future of Automated Workflows The integration of these agents aims to remove the friction of copying files between local software and cloud storage. By bringing the software to the data, Dropbox enables workflows like video editing and spreadsheet processing directly in the browser. This shift focuses on user intent rather than the technical setup of environments. ## Strategic Impact of Integration This development marks a shift toward agents that act as collaborative partners rather than just assistants. By embedding security and version history into the agent's core, Dropbox provides a reliable foundation for automation. The ability to trace every file change back to a specific conversation request is essential for maintaining user trust. ## Sources covering this story - https://dropbox.tech/machine-learning/how-two-engineer-side-projects-became-the-agent-taking-action-inside-dropbox — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._