How is SPACE different from other agent runtimes like Cursor Automations or Devin's cloud sandbox?

SPACE is designed for multi-hour to multi-day AI agent sessions with full VM isolation. Cursor Automations trigger short agent runs from codebase changes, Slack, or timers (best for CI-style tasks). Devin's cloud sandboxes are Docker-container-based and optimized for coding workflows. SPACE's differentiator is the rolling-snapshot resume model — an agent can pause overnight and resume with full state intact — plus stricter VM isolation for enterprise credential handling.

Can I use SPACE for my own agents, or is it Perplexity-only?

As of July 2026, SPACE powers Perplexity Computer sessions but hasn't been offered as a standalone runtime for third-party agents. Perplexity has published research (research.perplexity.ai) describing the architecture, and the company's positioning suggests eventual availability — but no public API or GA date has been announced. Third-party developers should track Perplexity's enterprise / API updates for a SPACE-as-a-service offering.

Why does SPACE matter for the AI agent market?

Long-running agents (runs measured in hours or days) have historically been unstable — timeout errors, state loss, credential leakage across tenants. SPACE is one of the first production-scale demonstrations that VM-isolated, snapshot-resumable agent runtimes work at real user load. It's a template other agent-runtime vendors (Cognition/Devin, Cursor, OpenAI's ChatGPT Work) are likely to copy or license as agent session durations extend from minutes to days.

Quick Answer

What Is Perplexity SPACE? Agent Sandbox (Jul 2026)

Q: What is Perplexity SPACE?

Perplexity SPACE is a secure, VM-isolated sandbox runtime for long-running AI agents, publicly unveiled on July 15, 2026. It has been powering all production traffic for Perplexity Computer (Perplexity's autonomous browsing / coding agent) since June 2026. Unlike traditional sandboxes designed for brief tasks, SPACE captures live memory snapshots roughly every minute so an agent can pause for a week and resume exactly where it left off. The architecture uses VM-based isolation, rolling snapshots, and external credential controls.

Published: July 19, 2026

What Is Perplexity SPACE? Agent Sandbox (Jul 2026)

On July 15, 2026, Perplexity publicly unveiled SPACE — a VM-isolated sandbox runtime designed specifically for long-running AI agent sessions. It had already been powering all production traffic for Perplexity Computer (Perplexity’s autonomous browsing / coding agent) since June 2026.

SPACE stands for Sandboxed Platform for Agentic Code Execution. The engineering headline: rolling live-memory snapshots roughly every minute, so an agent can pause for hours or days and resume in exactly the same VM state — not restarted, not re-planned, actually resumed.

Last verified: July 19, 2026

The Basics

Spec	Perplexity SPACE
Vendor	Perplexity AI
Public unveil date	July 15, 2026
Serving production traffic since	June 2026
Primary use	Runtime for Perplexity Computer agent sessions
Isolation model	VM-based (per-session VM)
State model	Rolling live-memory snapshots (~1-minute cadence)
Max session length	Multi-day (agent can resume after a week)
Credential handling	External credential broker; not in-VM
Third-party availability	Not GA as standalone runtime (as of July 2026)
Documentation	Perplexity Research: “Making SPACE: secure and efficient runtimes for long-running agents”

What Problem SPACE Solves

Traditional agent sandboxes (Docker containers, ephemeral cloud sandboxes) are designed for one-shot tasks: spin up, execute a script, tear down. They fail badly at:

Multi-hour and multi-day sessions. State drifts, memory leaks, container restarts kill the run.
Human-in-the-loop pauses. If a human needs to review at hour 3 and approve at hour 8, the container sitting idle is either killed or resource-wasteful.
Credential leakage across tenants. Container-level isolation is weaker than VM-level.
Deterministic resume. After a crash or scale-down, re-running from the beginning is expensive.

SPACE targets each:

Rolling snapshots. Every ~1 minute, SPACE captures live VM memory, disk, and network state. If the VM dies, resume from the last snapshot. If the human takes a lunch break, the VM parks efficiently.
Pause / resume. An agent can be paused for a week and resumed. State is preserved.
VM isolation. Full-VM per session; credentials handled by an external broker.
No re-plan on crash. The agent doesn’t have to re-reason about what it was doing — it wakes up mid-step.

The Architecture (What Perplexity Published)

From Perplexity Research’s SPACE paper (July 15, 2026):

Per-session VM. Each Perplexity Computer session gets a dedicated VM, not a shared container.
Snapshot cadence. ~1-minute rolling snapshots (live memory + disk state).
Credential broker. Credentials are held outside the VM by a hardened broker; the VM asks for them when needed and never persists them.
Network egress control. Fine-grained egress policy per session so agents can’t exfiltrate.
Resume semantics. Restore from any snapshot within retention; SPACE guarantees the resumed VM is bit-identical to the paused state for scheduled or ad-hoc resumption.

How SPACE Compares to Other Agent Runtimes

Runtime	Vendor	Isolation	Best For	Session Length
SPACE	Perplexity	Per-session VM	Long-running autonomous agents	Multi-day
Devin cloud sandbox	Cognition	Container	Coding agent workflows	Hours
Cursor Automations	Cursor	Cloud sandbox	Codebase / Slack triggers	Short (minutes-hours)
ChatGPT Work agent	OpenAI	Managed cloud	Office task automation	Session-based
Claude Cowork	Anthropic	Managed cloud	Multi-tab coding + web	Persistent web/mobile
E2B / Modal sandboxes	E2B, Modal	Container / Firecracker	Third-party agent devs	Task-scoped
AWS Bedrock Agents	AWS	Lambda / ECS	Enterprise agent runtime	Session-based

SPACE’s unique niche: VM-level isolation + minute-cadence snapshotting + pause-and-resume across days. No competitor combines all three at production scale as of July 2026.

Sub-Questions People Are Asking

Is SPACE open-source? No. Perplexity has published research describing the architecture but has not open-sourced the runtime.

Can I use SPACE outside Perplexity Computer? Not as of July 2026. It’s the runtime for Perplexity’s own agent product. Any third-party availability would be a future enterprise offering.

How does SPACE compare to Anthropic’s Computer Use environment? Anthropic’s Computer Use is a model capability (Claude can drive a browser and OS). It runs on whatever sandbox the developer provides. SPACE is that underlying sandbox — for Perplexity, not Anthropic. Anthropic developers still bring their own runtime (E2B, Modal, Docker, etc.).

Does SPACE make Perplexity Computer more expensive? It’s an infrastructure investment on Perplexity’s side. End-user pricing for Perplexity Computer hasn’t changed with SPACE’s unveiling (subscription-based, no per-VM billing announced).

How does snapshot-resume work with LLM state? The LLM itself (Sonnet 5, Opus 4.8, or whichever model powers the agent) is external to SPACE — the VM state is browser state, terminal state, file state, working memory. When resumed, the same LLM continues the conversation using the tool-call history and file state that SPACE preserved. Nothing about the LLM inference is snapshotted; the surrounding environment is.

What’s the security model for credentials? External broker. When an agent needs (e.g.) a Gmail token to compose an email, SPACE requests it from a hardened broker service that has been granted access by the user. The token never persists in the VM’s disk or memory beyond the immediate use. This is stronger than in-VM credential storage (common in Docker container agent setups).

How is SPACE billed? Not disclosed publicly. Internal to Perplexity for now.

Why SPACE Matters for the Agent Market

1. Sets a bar for enterprise agent runtimes. VM-level isolation + snapshot resume + external credentials is the strongest security posture for agent runtimes shipped to date. Enterprises evaluating Devin, Claude Cowork, or ChatGPT Work will now ask their vendors “how do you compare to SPACE?” — and Perplexity has published research they can point to.

2. Enables week-long agent tasks. Real research agents, engineering-refactor agents, sales-outreach agents all take days-to-weeks. SPACE proves it’s engineerable at production scale.

3. Puts pressure on Cognition and Cursor. Devin’s Docker sandboxes and Cursor’s cloud sandboxes are both simpler and lower-cost than SPACE — but for long, expensive agent runs, buyers may prefer SPACE-class runtimes. Expect Cognition and Cursor to publish or ship SPACE-equivalent architectures within 6-12 months.

4. Validates VM-per-session at scale. Historically people assumed VM-per-agent-session was too expensive at scale (VMs are heavier than containers). SPACE proves the economics work when snapshots enable efficient pause states. That’s a real engineering result.

5. Replaces Perplexity Computer’s Wikipedia-defined “vaporware” reputation. Perplexity Computer had been criticized as an unfinished / unreliable agent product through Q2 2026. Moving to SPACE (production traffic since June) explains why Computer stabilized and is a real answer to reliability critics.

The Competitive Landscape

SPACE vs Cursor Automations (July 11, 2026): Cursor Automations triggers short agent runs from codebase changes, Slack messages, or scheduled timers — cloud sandboxes optimized for CI-style workloads. SPACE is for autonomous long-running sessions that pause and resume. Different problem, different fit.

SPACE vs Devin’s cloud sandboxes: Devin’s coding sandboxes are container-based and optimized for coding-agent workflows (fast startup, integrated with Cognition’s agent stack). SPACE has stronger isolation and better long-session support. Devin can add snapshot-resume; SPACE would need to add coding-specific developer ergonomics.

SPACE vs ChatGPT Work agent (July 9, 2026, replaced Atlas): ChatGPT Work is a task-oriented agent (invoices, reports, HR ops) with its own managed cloud. Perplexity Computer + SPACE is a general-purpose autonomous computer-use agent. Overlap is limited today but growing.

SPACE vs Claude Cowork (July 7, 2026): Cowork is persistent conversation state across web / mobile / desktop, powered by Anthropic-managed cloud. Different problem: Cowork is about session continuity for the human-in-the-loop; SPACE is about VM continuity for the agent-alone.

The Question People Should Be Asking

When will Perplexity offer SPACE as a service?

If Perplexity monetizes SPACE as a hosted agent runtime for third-party developers, it becomes a direct competitor to E2B, Modal’s Sandboxes, Cloudflare’s Isolate-based agent runtimes, and AWS Bedrock Agents. That’s a large addressable market — probably larger than Perplexity’s search / answer business today.

Watch for a Perplexity API / Enterprise announcement in Q3-Q4 2026 offering “Perplexity SPACE as a runtime” pricing. If it lands at competitive per-session pricing versus Modal or E2B, it becomes an infrastructure-layer product that outlasts the current agent hype cycle.

Bottom Line

Perplexity SPACE is the first production-scale, VM-isolated, snapshot-resumable agent runtime. It’s currently limited to Perplexity Computer, but the published research and architecture set a new bar for what enterprise-grade agent runtimes should look like.

Who cares:

Perplexity Computer users — you now understand why the product is more reliable in July than in Q1.
Agent-runtime vendors (Cognition, Cursor, E2B, Modal, AWS) — you now have a competitive spec to match or beat.
Enterprises evaluating agent platforms — VM isolation + snapshot resume + external credential broker is the security posture to ask for.
Developers building third-party agents — watch for a Perplexity SPACE API. It could reshape the runtime market.

The short version: SPACE is Perplexity’s answer to “how do you run agents that don’t crash after 3 hours.” That’s a real problem, and their answer is worth studying.

Sources

Perplexity Research SPACE paper: research.perplexity.ai/articles/making-space-secure-and-efficient-runtimes-for-long-running-agents
Techstrong.ai SPACE coverage: techstrong.ai/articles/perplexity-unveils-space-a-secure-sandbox-platform-for-ai-agents
AI Business enterprise-AI context: aibusiness.com/agentic-ai/perplexity-ai-introduces-sandbox-for-agents
Financial Express weekly AI roundup July 19, 2026: financialexpress.com/life/technology-chinese-kimi-k3-outshines-fable-openais-codex-hardware-and-perplexity-returns-with-space-weekly-ai-roundup-19-july-4296162