Perplexity rebuilt its AI agent sandboxes around durability, not isolation

Perplexity launched SPACE on July 15, the sandbox platform now running underneath Comet, its agent platform for knowledge work. The company built it despite a maturing market of off-the-shelf options, including Firecracker (the open-source microVM technology AWS built for Lambda) and a growing roster of managed sandbox providers, because its infrastructure team concluded isolation wasn't the actual hard problem for running millions of concurrent agent sandboxes.

The harder problem, according to Perplexity infrastructure VP Nate Kupp, is durability. SPACE takes rolling snapshots of full session state, including live memory, as often as every minute, letting a session rewind up to a week. Kupp says that architecture change delivered "3x or more speedups on all of the metrics that we care about for performance," plus meaningful cost improvements over incumbent providers, with the team specifically tuning P95 and P99 tail latencies since sandbox performance sits directly in the user's critical path.

"Stateful systems are incredibly hard to build," Kupp says, pointing to the durability and uptime guarantees Perplexity has to hit while running millions of sandboxes at once. Rather than treating security and performance as a trade-off to balance, Kupp frames it as a knobs problem, giving customers the right controls rather than fixing one point on the spectrum for everyone.

For a small team building its own agent product, the takeaway isn't to build a custom sandbox platform, that's a scale problem most teams don't have. It's that the sandbox layer under an agent does more than isolation once sessions need to persist and resume, worth checking before assuming an off-the-shelf microVM setup covers it.