Perplexity open-sources SPACE: what a public agent sandbox means for business file workflows

What Perplexity announced

Perplexity's dated October 15, 2026 announcement says the company is open-sourcing SPACE, the in-house platform it uses to run agent code, and made the source repository publicly accessible the same day. The project's guide describes the platform as a sandbox for agent and user code that runs each session inside a separate Firecracker microVM, a lightweight virtual machine that keeps the code away from the host environment. The guide documents HTTP commands for creating a workspace, executing commands, reading output, and managing a session's lifecycle, including pausing, resuming, recovering, and branching a session. Isolation, recovery, and pause and resume are therefore documented features of the published code, not independent findings of this article.

Independent reporting corroborates the announcement. Semafor's October 15 article by Reed Albergotti describes the same open-source release and includes an interview attributed to Perplexity technical staff. The vendor's July description of SPACE as an internal deployment is a separate, earlier event; what is new on October 15 is that the code, the repository, and the setup guide are public.

The published setup guide also draws a clear operational boundary. Local deployment requires Linux on x86_64 hardware with KVM virtualization and a privileged Docker and Kind configuration, and the guide instructs operators to keep the development setup private because it is not a production configuration. An announcement and a public repository do not, by themselves, demonstrate production security, and Detailed In Design did not install, benchmark, or audit the software for this article.

What the release does not establish

The public code does not prove that any sandbox is escape-proof or that agent code will behave correctly inside it. Reporting and vendor statements about safety testing describe different scopes and different counts, and this article deliberately repeats neither figure nor any security-result claim drawn from them. A sandbox constrains where code runs; it cannot guarantee correct calculations, compliant handling of customer data, or immunity from attacks. Claims about performance multipliers, scale, or cloud availability are vendor-attributed and are not evaluated here.

What this means for your business

The transferable idea is the boundary, not the vendor. Spreadsheet and file-processing assistants need a bounded place to work, recoverable files, and a clear staff approval step before business records change. Detailed In Design could build a staff-facing supplier-spreadsheet intake and reconciliation portal around that pattern: a staff member uploads an authorized copy of a supplier file, cleaning and reconciliation run in a separately managed execution workspace rather than against live records, the portal displays proposed changes with source references so staff can see exactly what will change and why, and an explicit staff approval is required before the system exports results or writes to an authorized ERP or CRM system. SPACE is one runtime option we could evaluate for that isolated workspace, not a promised component, and this proposed workflow is separate from Perplexity's release; it is not a vendor partnership, an existing Detailed In Design feature, or an already-integrated system.

The expected benefit is procedural rather than numerical. Isolated processing with visible diffs and a required approval step makes data preparation more inspectable, reduces ad hoc file handoffs, and leaves a clearer record of who approved what. We are not claiming measured savings, accuracy gains, or error reductions, because none have been measured for this proposed workflow.

The dependencies and limits are real. A project like this requires authorized copies of source files, approved ERP or CRM interfaces, defined hosting and operations scope, representative tests with your own data, credential and network policy, data retention rules, named human reviewers, and clear rollback and export rules. A sandbox cannot ensure correct calculations, compliant data handling, or protection from every attack, and the approval step is only as good as the reviewers who act on it.

A practical next step

If your team already uses AI assistants on spreadsheets or documents, a bounded pilot is a reasonable first move: one authorized file set, one isolated workspace, one visible diff report, and one named approver, with success judged on whether staff can see and verify every proposed change before it reaches a system of record. Detailed In Design can discuss scoping such a pilot using the custom software, AI integration, and document workflow capabilities described on our site.