English for Vercel Sandbox

Running untrusted or dynamically generated code — often produced by an AI agent — safely is a problem more teams face every year, and Vercel Sandbox is one of the tools built specifically for it. Talking about sandboxed execution requires vocabulary that’s a little different from typical serverless discussions: “ephemeral compute,” “isolation,” and “sandbox lifecycle” all describe distinct guarantees. This guide covers the terms you’ll need.

Key Vocabulary

Ephemeral compute — a compute environment that’s created for a single task and destroyed immediately afterward, leaving no persistent state behind between runs. “Because it’s ephemeral compute, we don’t have to worry about one user’s generated code leaving artifacts that affect the next execution.”

Isolated execution — running code in an environment that’s strictly separated from the host system and other running sandboxes, so a failure or malicious action in one can’t affect others. “Isolated execution is the whole point here — even if the generated code tries something malicious, it can’t reach our actual infrastructure.”

Sandbox lifecycle — the sequence of states a sandbox moves through: creation, execution, and teardown, typically within a bounded time limit. “We’re seeing timeouts because the sandbox lifecycle has a hard cap, and this particular workload runs longer than that.”

Untrusted code execution — running code that hasn’t been reviewed or vetted by your team, commonly generated dynamically by a user or an AI agent, which requires stronger isolation guarantees than your own application code. “This is untrusted code execution, so it goes through the sandbox — we never run agent-generated scripts directly on our own servers.”

Filesystem snapshot — a captured state of a sandbox’s filesystem, either used to initialize a new sandbox quickly or to inspect what a completed sandbox produced. “We take a filesystem snapshot after the sandbox finishes so we can retrieve the generated files without keeping the sandbox itself running.”

Resource limits — constraints placed on a sandbox’s CPU, memory, and execution time, preventing a single runaway process from consuming excessive resources or running indefinitely. “The generated script hit our resource limits and was killed automatically — that’s expected behavior, not a bug.”

Common Phrases

• “This has to run in a sandbox — it’s untrusted code execution, not something we wrote ourselves.”
• “The sandbox lifecycle timed out before the build finished; we may need a longer execution window for this workload.”
• “Grab the filesystem snapshot before the sandbox tears down, or we lose the output.”
• “That’s isolated execution working as intended — the failing process couldn’t touch anything outside its own sandbox.”
• “We hit resource limits on that run; let’s check whether the generated code has a runaway loop.”

Example Sentences

Explaining the architecture to a security reviewer: “Any code generated by the AI agent runs inside an isolated sandbox with strict resource limits and a bounded lifecycle — it never has direct access to our production systems, and the entire environment is destroyed once execution finishes.”

Reporting an issue: “Several sandbox executions are hitting the timeout before completing. It looks like this particular workload — installing dependencies and running a build — needs more time than our current sandbox lifecycle allows.”

Describing the design rationale in a planning doc: “Because we’re letting an AI agent generate and execute arbitrary code on a user’s behalf, we can’t run that in our own application process. Ephemeral, isolated sandboxes let us support that use case without exposing the rest of our infrastructure to untrusted code.”

Professional Tips

• Use “untrusted code execution” explicitly when justifying a sandboxing decision to a security-conscious audience — it names the actual risk being mitigated, rather than leaving it implied.
• Distinguish ephemeral compute (no persistent state) from isolated execution (no cross-contamination) — they’re both properties of a sandbox but address different concerns.
• When a workload fails, check whether it’s a sandbox lifecycle timeout or a resource limit being hit — the fix differs (extend the time budget vs. optimize the workload).
• Mention filesystem snapshots specifically when describing how you retrieve output from a sandbox that’s about to be torn down.

Practice Exercise

1. Explain, in two sentences, why AI-agent-generated code should run in a sandbox rather than your main application process.
2. Write a one-sentence bug report describing a sandbox timing out before finishing a task.
3. Describe the difference between ephemeral compute and isolated execution to a teammate unfamiliar with sandboxing.