Updates·July 15, 2026, 22:48

Perplexity builds protected playground for long-running AI agents

AI-generated from the sources listed, with automated source checks.

Perplexity has described SPACE, a platform that gives AI agents an isolated environment to work in. The speed figures are the company's own, and an independent analysis suggests that published details are lacking.

AI-generated image

Perplexity published a technical article on SPACE on July 15, 2026. It is a so-called sandbox platform for AI agents that are meant to run for long periods. A sandbox is a closed area where a program can work freely without damaging anything outside it.

How the sandbox is protected

According to Perplexity, an agent in SPACE can work freely while the host (the computer that runs the sandbox), other users and sensitive login credentials are protected.

The protection is built in several layers: a virtual machine for each sandbox, isolation of processes on the host system, controlled communication channels and a network gateway that governs what is sent out.

Secret keys and similar login credentials never enter the sandbox itself. Instead, they are injected at the network layer. For ordinary users, this means an agent can use a service without ever seeing the password for it.

Faster startup, according to Perplexity

Perplexity says the median time to create a sandbox fell from 185 to 60 milliseconds, or 3.1 times faster. For the slowest tenth, the time fell from 447 to 89 milliseconds, 5.0 times faster. The figures are Perplexity's own.

During the rollout, SPACE handled millions of created sandboxes and tens of millions of reconnections, according to the company.

SPACE currently powers Perplexity's internal agent product Computer. The article indicates that the platform could later be extended to other platforms.

Earlier steps and a critical look

According to GadgetBond's coverage, Perplexity already announced a Sandbox API for developers on March 12, 2026, allowing AI agents to run code in isolated environments. It was described as upcoming and in private beta. It ran in separate Kubernetes pods with a persistent file system, and sessions could be paused and resumed hours later.

An independent analysis from The Model Wire sees the launch as a shift from search toward agent infrastructure and as competition with OpenAI's ecosystem. At the same time, the analysis notes that Perplexity has not published benchmarks, latency figures or specifications for tool integration. According to The Model Wire, the launch therefore looks like positioning ahead of a finished product.

Sources

More on this topic

Get the week's AI news in your inbox

Choose your level, topics and length. One email a week, unsubscribe at any time.

Subscribe to Promptly Newsletter