# Run Scale AI’s AgentEnv evals on Vercel Sandbox

Source: [Vercel](https://vercel.com/changelog/run-scale-ais-agentenv-evals-on-vercel-sandbox)  
Feed7 permalink: https://feed7.dev/p/run-scale-ais-agentenv-evals-on-vercel-sandbox-0cdbcpd  
Published: 2026-10-26T19:00:00.000Z  
Trust: Official Source (official_source)

## Why Included

AgentEnv 0.9.1334 and later can run evaluations on Vercel Sandbox, keeping the agent and task environment in separate sandboxes. Builders can inspect task scores and recorded actions.

## Source Summary

**AgentEnv 0.9.1334 and later** includes a Vercel provider. It runs the agent and task environment in **separate Sandboxes**; the example exposes task scores and recorded actions.

## Practical Implication

Builders evaluating agents can use this setup to run tasks in isolated environments and inspect both the result and the agent’s actions.

## Agent-Ready Context

**AgentEnv 0.9.1334 and later** includes a Vercel provider. It runs the agent and task environment in **separate Sandboxes**; the example exposes task scores and recorded actions.

Builders evaluating agents can use this setup to run tasks in isolated environments and inspect both the result and the agent’s actions.

The example uses an existing Vercel CLI login and project. The material does not specify resource limits, pricing, or how scores are calculated.

## Connected Context

Feed7 judgment across 928 accumulated Signals:

This provides a concrete way to run AgentEnv tasks with the agent and task environment in separate Sandboxes, then inspect scores and recorded actions. It supports isolated evaluation, while leaving score methodology and resource costs unclear.

- [Run Terminal-Bench and other Harbor evals on Vercel Sandbox](https://feed7.dev/p/run-terminal-bench-and-other-harbor-evals-on-vercel-sandbox-0o8aa4r) — Both use Vercel Sandboxes for agent evaluation; Harbor adds per-trial microVM isolation and parallel runs, while AgentEnv specifies separate Sandboxes for the agent and task environment.
- [Quantifying infrastructure noise in agentic coding evals](https://feed7.dev/p/infrastructure-noise-1jyyyw1) — AgentEnv exposes task scores but does not specify resource limits; Anthropic’s findings show why documenting environment settings matters when interpreting score differences.
- [Everything Is a Rollout — Alex Shaw + Ryan Marten, Terminal-Bench, Harbor, Laude Institute](https://feed7.dev/p/everything-is-a-rollout-alex-shaw-ryan-marten-terminal-bench-harbor-laud-0iz4rgx) — Harbor’s emphasis on reproducible runs and inspecting agent actions reinforces the value of AgentEnv’s recorded actions alongside task scores.

## Context Map

- Layer: benchmark
- Domains: coding
- Topics: agent-evals, sandboxing

## Uncertainty

- The example uses an existing Vercel CLI login and project. The material does not specify resource limits, pricing, or how scores are calculated.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.