Key Takeaways

  • Modal sandboxes started as an underused side project in 2023, but they now generate one third of the company's total revenue.
  • Sandbox revenue continues to expand by 30% month over month, defying Bernhardsson's repeated expectations of a plateau.
  • The median Modal sandbox lives for just eight seconds, built for fast single-command execution rather than continuous uptime.
  • Treating agent sandboxes as long-running remote developer boxes is a mistaken lift-and-shift mental model that inflates costs and latency.
  • True agent infrastructure requires sub-second spin-up times and container snapshotting so developers can pause, resume, and clone state instantly.

The Lift-and-Shift Trap of Remote Dev Boxes

Most teams building coding agents copy the human developer workflow. They spin up a virtual machine, install an IDE environment, and leave it running for hours. Bernhardsson argues this approach copies the wrong pattern: “A lot of long running persistent sandbox stuff right now today feels lift and shift. I suspect sandboxes will evolve from dev box in the cloud to something more ephemeral. One environment for every command and snapshotting in between so you can pause resume clone.”

A human sits at a terminal, edits files, and leaves background services running. An autonomous agent works in bursts. It inspects a file tree, runs a linter, executes a bash script, and halts until the next inference step. If you pay for an idle cloud workstation between those steps, you pay for dead time. Bernhardsson watched this shift happen directly inside Modal. What looked like a niche experiment in 2023 quickly exploded into one third of their revenue. As Bernhardsson put it: “Every month I keep saying our sandbox revenue might have plateaued but it keeps going up 30% every month.”

Why Speed Dictates Agent Architecture

The real bottleneck in running autonomous coding agents is startup overhead. If a new container takes fifteen seconds to boot, you have no choice but to keep a persistent box alive. When startup latency drops below a single second, the entire operational model changes.

“I think the median sandbox in Modal today is like 8 seconds long or something like that, which is kind of incredible when you think about it, and over time I think it might get even shorter,” Bernhardsson noted. “The faster you can get that wake up, go to sleep cycle, the more you can create sandboxes that only run for a very brief period. And that will make compute much more efficient.”

This short lifespan changes what you can build. Instead of running a single agent sequentially inside one container, you can split execution into independent environments. Bernhardsson pointed to test suites as the cleanest example: “What I would love to do is take your entire test suite and just run every single test in its own sandbox, and just parallelize that around like a thousand sandboxes. Why can't that be done? The problem is there's too much overhead around starting and stopping sandboxes.”

Once starting and stopping takes milliseconds, durability moves to the orchestration layer. Your orchestrator tracks the agent state, while the sandbox acts as an expendable execution shell that dies the moment a command finishes.

What to Do With This

Audit your agent pipeline this week and measure the idle minutes your execution environments spend waiting for model responses. Replace your persistent virtual machines with ephemeral container calls that spin up per action and snapshot modified state back to object storage. If an agent task runs sequentially over five minutes, break it into parallel sandbox jobs that run concurrently in ten seconds.