Large language models (LLMs) are reaching impressive levels of capability when it comes to planning and carrying out complex tasks that stretch across many steps and significant amounts of time.
Yet this potential often stays theoretical unless the models have access to a secure, persistent place where they can actually run code, manage files, and interact with tools without putting the rest of the system at risk.
Traditional approaches to running code in isolated environments were built for quick, one off jobs that do not carry state forward. They typically share a single kernel across workloads, which creates a weak point if something inside behaves unexpectedly or maliciously.
For agents that need to accumulate context, keep a working filesystem alive, and resume after pauses that might last minutes or hours, these older methods simply do not provide enough isolation or continuity.
Perplexity developed SPACE to close that gap.
According to Perplexity in a post on its website, SPACE, or short for Sandboxed Platform for Agentic Code Execution, gives long running agents a purpose built home that balances strong protection with the practical features they need to operate effectively at scale.
It has already become the foundation for Perplexity Computer, handling every session in that product and proving itself under real production load.
The system uses three coordinated layers.
A control plane manages overall state and requests. Node level services handle lifecycle, storage, and safety rules.
Each sandbox runs as its own virtual machine with a dedicated guest kernel and a space daemon that safely brokers interactions with the host. This setup contains any issues within clear hardware boundaries.
Security is enforced through multiple layers including external credential management, controlled network egress, and encryption options.
Functionality supports a rich lifecycle with fast creation, pausing, resuming, and snapshots for rollback and recovery. Efficiency comes from the btrfs filesystem, which enables quick copy on write clones and space saving snapshots along with warm template pools.
The result is sandbox creation that is three to five times faster than before.
The combination of these choices delivers something greater than the sum of its parts.
Agents gain a reliable substrate where they can pursue ambitious goals without constant restarts or fear of side effects. Sessions can branch, recover, and persist across long periods while the platform keeps the host, other tenants, and user secrets protected by default. Perplexity has already seen the difference in practice as SPACE supports the demanding workloads inside Computer.
Looking forward, the same foundation is being prepared for broader use.
Plans include extending SPACE to additional guest types and deployment settings so developers can bring secure, stateful agent runtimes to more places where people actually work.
As large language models continue to grow stronger at reasoning and tool use, infrastructure like this turns that strength into dependable, long running action. It gives agents the room they need to operate while keeping everything else safe, efficient, and ready for whatever comes next.
