All posts
Company
7 min read
What we learned running 10,000 agents for a week
We stress‑tested the platform with a fleet of browsing agents. Some numbers, some surprises, and the changes we shipped because of it.
The numbers
Ten thousand machines, seven days, roughly 1.4 million browser sessions. Median wake latency stayed under 50 ms throughout. Disk churn was the surprise: agents write far more than we expected, mostly caches and screenshots, so we shipped automatic cache pruning and cheaper snapshot storage.
The other lesson was about sleep. Machines that sleep aggressively cost almost nothing, and wake fast enough that users never notice. Sleep is now the default after ninety seconds idle.
