Modal Can Spin Up a Million AI Sandboxes in Under a Minute
They rebuilt the platform to handle huge numbers of agents without the usual slowdowns
@modalTL;DR:
- Modal's update lets them create tens of thousands of sandboxes per second
- They ditched central schedulers and slow databases to speed things up
- This targets AI agents, RL, and bursty workloads that need quick isolation
Headline
Modal says it can now launch a million AI agent sandboxes in under a minute.
Summary
The company reworked its sandbox platform so it can run millions of them at once and create tens of thousands every second. The goal is to handle reinforcement learning jobs, background agents, and sudden spikes in agent traffic. They cut out central coordination and slow shared databases from the creation path. Instead they use scheduling servers that scale out and keep most state right on the workers.
Analysis
This matters for anyone running lots of AI agents. Coding agents, browser agents, RL rollouts, and automated tasks all need fast, isolated environments. Modal points out that Kubernetes-style setups weren't built for this kind of churn and concurrency. By giving workers more control and handling metadata asynchronously, they traded some consistency for speed. If it holds up in real use, it could make Modal a stronger choice for agent companies that need to spin up huge numbers of sandboxes on demand.
Impact Assessment
- High significance for developer tools and AI infrastructure
- Relevant to industry trends around agent platforms and technical execution backends