Perplexity launched Sandbox API, a standalone service that provides isolated code execution environments for AI agents requiring deterministic computation.

The API emerged from Perplexity's internal infrastructure. The company's Computer product, Finance Agent, and Deep Research tools all execute code inside sandboxed environments, with Computer running thousands of sessions per minute.

Stateful execution with persistent storage

Each session runs in an isolated Kubernetes pod with a persistent filesystem mounted via FUSE. The FUSE daemon intercepts file operations and translates them for agents to read, write, list files, and track modifications.

Sessions maintain state across workflow steps. Files created in one execution remain available in subsequent steps. Long-running workflows can pause and resume hours later with full state intact.

The service supports Python, JavaScript, and SQL with runtime package installation per session. Each session can run up to five background processes simultaneously.

Zero-trust security architecture

Sandboxes operate with no direct network access by default. When outbound connectivity is required, traffic routes through an egress proxy running outside the sandbox environment.

The proxy matches outbound requests by destination domain and injects appropriate credentials. Code executing inside sandboxes never accesses raw API keys or secrets directly.

Built-in timeouts and resource limits enforce execution boundaries across all sessions.

Integration with Agent API

Sandbox API will integrate as a tool within Perplexity's Agent API, allowing the orchestration runtime to delegate deterministic code execution mid-workflow. Agents decide what to compute, dispatch to Sandbox, observe outputs, and continue reasoning.

The integration uses the same API key and credit system as existing Perplexity services.

Perplexity will open a private beta shortly, with documentation available at docs.perplexity.ai.