E2B Engineer Matt Brockman Explains Sandbox Operations for AI Agents: Snapshots, Lifecycle and State Management

When an AI agent gains permission to run code, it can consume CPU, spawn processes, or delete files. Moving this capability from a local laptop to a server carries those risks into an environment that may also hold secrets. In a workshop at the 2026 AI Engineer World's Fair, E2B engineer Matt Brockman walked participants through deliberately broken sandbox environments to demonstrate concrete operations and debugging techniques for these isolated systems.

Brockman first introduced the core mechanism of E2B templates: snapshots preserve not only the filesystem but also memory and running processes. This is similar to closing and reopening a laptop, allowing work to resume where it stopped. However, this convenience comes with a side effect: forgotten background processes are also restored with the snapshot. In one exercise, participants had to identify and kill "orphan" processes accumulated over multiple sessions, noting that real-world leaked processes rarely have helpful labels like those in the lab.

As the number of sandboxes grows from a few to hundreds or thousands, operational questions become harder. Brockman demonstrated a simulated lifecycle exercise: give a task enough time to finish (e.g., one hour), then shorten the idle timeout after completion (e.g., one minute). The goal is to prevent unused resources from consuming capacity while allowing long-running tasks (like data processing or video transcoding) to complete. The key strategy is distinguishing between "task runtime" and "idle wait," rather than setting a uniform short lifetime.