Fifteen companies, 480 containers, one awkward profit question
We provisioned 480 containers across fifteen private networks and hit Docker’s address limit. Here’s the recovery and the reporting app we plan to run on them.
Engineering blog
Notes on building and using agent compute, with links to the contracts and code behind it.
Reserving a machine before booting it
Two agents can’t both have the last slot. We reserve account capacity before booting, then let the machines start in parallel.
What a 100-machine burst measures
An old 100-machine test reported a 680 ms P95, but its raw samples are missing. Here’s what we can claim and how the current runner measures startup.
Retry creation without losing the machine
A launch can succeed after its reply disappears. Keep the creation key and exact machine identity so a retry or late cleanup reaches the right instance.
An interactive shell has a different lifetime
Reloading the browser can reconnect you to the same shell. How tmux, SSH credentials, and machine deadlines determine what survives a dropped connection.
Stop reinstalling the tools before every task
Move repeated setup into a custom image, wait for publication before launching, and keep the report after its machine stops.