The workspace, not the chat, is the unit of work.
We built a collaborative AI application builder where teams — not individuals — are the primary unit of work. Multiple teammates share a workspace, watch the same agent build in real time, and the database itself prevents two runs from ever colliding.
The result: a structurally safe, cost-efficient platform where team collaboration is a data-model primitive, not a front-end feature.
Every AI builder was built for one person with one prompt box.
The moment a second teammate joined a project, existing tools broke down — there was no shared awareness of what the AI was doing, no role system controlling who could trigger a build, and no protection against two people firing conflicting agent runs on the same codebase simultaneously.
We treated this as a data-modeling problem before a UI problem. If 'team' and 'build' aren't first-class objects in the schema, no amount of front-end polish makes collaboration feel real.
Workspaces. Roles. One active run. By design, not by policy.
Workspace-First Data Model
Every project belongs to a workspace, and every build run belongs to a project — not to the user who typed the prompt. Team and role are schema primitives, not UI overlays.
Leased, Auto-Expiring Sandboxes
Cloud sandboxes are provisioned only when genuinely needed and carry their own expiry — idle projects are reclaimed automatically, so storage scales with active work, not total project count.
One Active Run Constraint
A unique database constraint makes two concurrent agent runs on one project structurally impossible — not just discouraged. The second request cannot start until the first completes.
Real-Time via DB Replication
The live build stream rides on the database's own real-time replication — no separate pub/sub service to manage, monitor or pay for.
Append-Only Event Log
Every build step is claimed by exactly one worker with a lease and retry ceiling. The history a team sees is the history that actually happened — no silent reruns, no collisions.
68% less storage. 71% lower cost. Zero concurrent-build incidents.
Storage Footprint per Project
From ~2.1 GB always-on to ~680 MB leased — storage scales with active work, not the total number of projects ever created.
Infrastructure Cost per Project
From $214 to $61 per active project per month — after the single-active-run constraint shipped.
Concurrent-Build Incidents
Across ~1,200 projects in four months — from 31 per 1,000 builds to zero after the database constraint went live.