“I love using Agent 37. I save like $2K a month, switched from AWS and Hetzner. Your tech is so reliable and efficient. We haven’t had any CPU problems since, and I’ve been coding on Vision and OGTool, multiple instances.”
The problem
Maddie Wang built Vision so a whole team can work on one production app at the same time. You connect your GitHub repo, everybody prompts in the same live workspace, and a running preview of the app sits next to the chat. She runs OGTool, her AI search visibility product, on the same stack.
Every one of those sessions is a real machine doing real work: installing dependencies, compiling, serving a preview. They all shared one 32 GB runner box. The memory limits added up faster than the box ever could.
“Sometimes apps are really RAM hungry. OGTool eats up like 10 GB, a huge spike, just while it compiles.”
When the box filled up, the node stopped heartbeating and every session running on it was failed. Eleven were on it the last time it happened. On the other side of that was a customer sitting in front of a workspace that would not start, and it landed hardest on the busiest days.
“So many dreams about CPU exploding.”
What they run now
Every session gets its own instance. Vision asks for one when somebody opens a workspace, picks the size that suits the app, and that session runs there alone with its own RAM ceiling. A heavy build hits its own wall and nobody else can feel it.
Vision hands the heavy apps 16 GB and 8 vCPU, so the ones that spike have the headroom in advance. OGTool’s 10 GB compile has room to spare on the size it gets, and the light apps are sized down and priced accordingly.
Vision turns auto-sleep on for every one of them, so an instance bills by the minute while it is awake and falls back to its disk alone once nobody is in it. Vision runs about 100 session-hours on an ordinary day and roughly twice that at peak, and pays for the hours people were actually building.
Vision and OGTool moved first. The rest of the workspaces followed over the next two days, with every batch watched before the next one went. Every workspace starts on Agent 37 now, and the AWS runner boxes it used to keep up for peak sit empty.
“I’m down to migrate the rest of the companies, today was so good for development.”
The outcome
- A heavy build stops at its own wall: Every session holds its own RAM ceiling, so a 10 GB compile is one session's problem and nobody else sits in a workspace that will not start.
- Sized per app, 16 GB for the heavy ones: Vision picks the size when it asks for the instance. The apps that spike get the room up front, and the small ones cost close to nothing.
- No fleet kept up for the peak: Vision runs every session with auto-sleep on, so an idle one pays for its disk alone and the bill follows the hours people actually spent building. Maddie puts the saving at about $2K a month.
Vision kept the part its customers pay for, which is a whole team building on one live app. The machine under each session stopped being something Maddie has to dream about.
See how instances are sized