Stacks/Knowledge work
Dropbox's file-acting agent
Dropbox merged two engineers' side projects into a customer agent: each conversation gets a fresh Ubuntu VM restored from a QEMU snapshot, the agent uses command-line software to work on files, and credentials and permission checks stay outside the VM. Plans, Allow/Deny cards, and version history keep changes reviewable and reversible; 75% of workspaces were ready within 300 ms in testing.
Daniel Reiter HornPublished Nov 5, 2026Last verified Nov 13, 2026
- Use case
- An agent inside Dropbox that answers questions about a user's files and carries out tasks such as organizing folders, renaming files, and preparing materials to share
- Category
- Knowledge work
- Team size
- not stated
- Orchestrator / framework
- The app sends each turn through a trusted host layer to the agent in a sandbox; broad requests must get a plan accepted first, progress shows as a checklist, and file changes require an Allow/Deny folder-access card. An adapter translates Pi harness messages into the Codex protocol the interface already speaks.
- Memory
- not stated
- Evals
- not stated
- Observability
- Each revision traces back to its conversation through Dropbox version history, which also supports undo.
- Cost (stated in writeup)
- Not stated
- Cost per output (stated)
- Not stated in the writeup
- WhatWeRun estimate
WhatWeRun estimate
Not enough info to estimate
The post doesn't name the model or give tokens or volume per task.
Reviewed Nov 13, 2026
- Observability tools
- not stated
- Other tools
- Ubuntu VM per conversation, restored from a QEMU memory snapshot (Riviera)rcloneDocument-conversion and image-processing CLI toolsAccess checks outside the VM; no general internet access
- Source
- https://dropbox.tech/machine-learning/how-two-engineer-side-projects-became-the-agent-taking-action-inside-dropbox
Daniel Reiter Horn · Published Nov 5, 2026 · Last verified Nov 13, 2026
Related
