Back to the stories

Perplexity Open-Sources SPACE Agent Sandbox

Score 10.0,

Artificial intelligence

Running autonomous agents inside a production-grade, safety-minded sandbox is no longer something you have to buy from a vendor.

Perplexity has open-sourced SPACE, the production-tested sandbox that runs behind its Computer product. The company published the code, setup and security guides on October 15, 2026, and released it under a permissive Apache 2.0 licence so teams can inspect, modify, and deploy it themselves.

Why this matters: agent runtimes need more than a model. They need isolation, credential controls, network gating and ways to recover state when something goes wrong. SPACE packages those operational pieces into a single, tested platform, so teams don’t have to assemble them from scratch or remain dependent on a commercial sandbox provider.

How it works in one image: each agent runs in its own tiny virtual environment (SPACE uses Firecracker microVMs), with APIs to create those environments, run commands, and control what the agent can access. Think of it as giving every agent a locked workshop, with a recovery plan for the workbench and separate keys for credentials and network access.

What changes now: developers and safety teams can run, audit and adapt a production-grade sandbox on their own infrastructure. Security practices that used to live behind closed doors are now something organizations can copy, test and improve. That said, open-sourcing the platform removes vendor lock but not the need for serious engineering and operations to run it safely.

What’s next: the real question is whether SPACE becomes the default safety baseline for agent deployments, or just one of many forks. Will independent audits and real-world rollouts turn it into common practice? We’ll watch how quickly teams adopt and harden it.