AI

Perplexity launches SPACE sandbox platform powering all Computer agent runtimes

Wednesday, July 15, 2026Read Original

Details

  • Perplexity introduces SPACE (Sandboxed Platform for Agentic Code Execution), the sandbox infrastructure behind Perplexity Computer.
  • SPACE creates isolated environments for code execution, file access, and long-running agent sessions, designed for workflows that run for hours or days.
  • The architecture separates the agent session from the sandbox: each task runs in a disposable Firecracker microVM that is destroyed once the work completes.
  • Rolling snapshots preserve live memory and files, enabling sessions to pause, resume, or branch across multiple sandboxes without persisting credentials inside the VM.
  • SPACE has handled 100% of Perplexity Computer’s production traffic since June, replacing earlier third-party sandbox infrastructure.
  • On identical production traffic, SPACE cut median sandbox creation latency from 185 ms to 60 ms and reduced P90 latency from 447 ms to 89 ms.
  • The platform is built for secure, efficient agent runtimes, keeping credentials outside the sandbox and using ephemeral microVMs to minimize blast radius and cross-tenant risk.
  • In its first weeks, SPACE has already processed millions of sandbox creations and tens of millions of reconnects for Perplexity Computer, demonstrating scalability under real-world load.
  • The system’s session/sandbox separation and rolling snapshot mechanism allow long-running tools and agents to maintain state over time while each individual sandbox remains short-lived and disposable.

Impact

By moving all Computer traffic onto SPACE, Perplexity is vertically integrating a key layer of agent infrastructure, trading dependency on generic short-lived sandboxes for a platform tuned to long-running, tool-using AI agents. This narrows the gap with rival AI stacks that are building proprietary runtime layers, and positions Perplexity to optimize cost, latency, and security as agent workloads scale.

Rift Dispatch