DEVIN
Cloud AI Workspace for Teams
A cloud workspace where software teams work alongside autonomous AI coding agents.
The product I'm building
Working mockup of the workspace. The interface will continue to change as the architecture develops.
WORKING MOCKUP · NOT FINAL UIThe Idea
A team creates a task, assigns humans and agents, gives agents a secure workspace, and reviews the resulting work.
How I'm designing it
The detailed architecture is still evolving. The Tldraw board shows the current design exploration.
VIEW FULL ARCHITECTUREWhat I'm solving right now
There will be more problems as the system grows. These are the ones I'm working through now.
How should the backend communicate with a long-running agent inside an isolated sandbox while keeping the frontend updated?
How should humans and multiple AI agents work around the same task while keeping context, ownership, and permissions clear?
What should the agent's cloud computer provide, and where do the sandbox's limits affect the architecture?
Problem → Exploration → Current Direction
Thinking through the system
System Layers
┌──────────────────────────────┐
│ CONTROL PLANE │
│ Organizations · Tasks │
│ Permissions · State │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ ORCHESTRATION │
│ Jobs · Agent lifecycle │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ AGENT RUNTIME │
│ Planning · Tools · State │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ SANDBOX │
│ Filesystem · Terminal · Git │
└──────────────────────────────┘The separation keeps coordination outside the environment where untrusted code actually runs.
Design Decisions in Progress
Connection model
SSE vs WebSockets
Agent location
Outside sandbox vs Inside sandbox
Execution coordination
Request lifecycle vs Long-running job