---
id: "nr-vercel-sandbox-now-runs-harbor-evals-including-terminal-bench-db3af1"
slug: "vercel-sandbox-now-runs-harbor-evals-including-terminal-bench-db3af1"
title: "Vercel Sandbox now runs Harbor evals including Terminal-Bench"
observed_at: "2026-09-18T13:48:49+00:00"
record_date: "2026-09-18"
date_kind: "observed_at"
novelty: "new"
source_platform: "x.com"
topics: ["evals","sandboxing","coding-agents"]
entities: ["Vercel Sandbox","Harbor","Terminal-Bench"]
related_patterns: []
source_url: "https://x.com/vercel_dev/status/2100649581482807686"
source_urls: ["https://x.com/vercel_dev/status/2100649581482807686"]
basket: {"id":"0325e6b8-f527-4f9b-bfa3-3f30e57f380b","revision":1,"cluster_id":"db3af174-4608-4e55-bdbb-837cdf5b3040"}
type: "post"
description: "Vercel says Harbor evals, including Terminal-Bench and more than 300 datasets, now run on Vercel Sandbox with each trial isolated in a microVM."
why_it_matters: "The post does not give pricing, concurrency limits, trial startup times or any benchmark results produced on the new setup. Published eval runs on Vercel Sandbox with reported costs and run times would show whether the setup is practical at scale."
verification_level: "source-inspected"
signal_type: "field_note"
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:vercel-sandbox-now-runs-harbor-evals-including-terminal-bench-db3af1"
retrieval_nugget: "Vercel says Harbor evals, including Terminal-Bench and more than 300 datasets, now run on Vercel Sandbox with each trial isolated in a microVM. Vercel announced on September 17, 2026 that Harbor now runs evals on Vercel Sandbox. According to the post, users can run Terminal-Bench or more than 300 other datasets, isolate each trial in its own microVM, keep credentials"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:576f21f32dab022aeeacae40c78577cfd6ce4d740809df3f921f410a4557c6d6","reviewed_at":"2026-09-24T13:28:46Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"The announcement treats agent evaluation as an isolation problem as much as a scoring problem.","observed_facts":[{"text":"Harbor now runs evals on Vercel Sandbox, according to Vercel Developers.","source_urls":["https://x.com/vercel_dev/status/2100649581482807686"]},{"text":"Each trial is isolated in its own microVM and credentials are kept out of the sandbox.","source_urls":["https://x.com/vercel_dev/status/2100649581482807686"]},{"text":"Models can be swapped with one AI Gateway key.","source_urls":["https://x.com/vercel_dev/status/2100649581482807686"]}],"mechanism":"Running each trial in its own microVM keeps state from one attempt out of the next, while holding credentials outside the sandbox limits what an agent under test can reach.","why_now":"This matters now because terminal-style benchmarks let agents execute arbitrary commands, which makes shared or credential-bearing environments a real risk during evaluation.","implications":["Teams running Terminal-Bench style evaluations can hand isolation to managed microVMs instead of maintaining their own execution environments."],"evidence_boundary":"The post does not give pricing, concurrency limits, trial startup times or any benchmark results produced on the new setup.","watch_conditions":["Published eval runs on Vercel Sandbox with reported costs and run times would show whether the setup is practical at scale."],"related_records":[],"new_branch_reason":"It connects a public agent benchmark harness to managed microVM isolation, a separate thread from model releases."}
routes: {"html":"https://newruntime.com/posts/vercel-sandbox-now-runs-harbor-evals-including-terminal-bench-db3af1/","markdown":"https://newruntime.com/posts/vercel-sandbox-now-runs-harbor-evals-including-terminal-bench-db3af1.md","json":"https://newruntime.com/posts/vercel-sandbox-now-runs-harbor-evals-including-terminal-bench-db3af1.json"}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
---

# Vercel Sandbox now runs Harbor evals including Terminal-Bench

## Retrieval answer

Vercel says Harbor evals, including Terminal-Bench and more than 300 datasets, now run on Vercel Sandbox with each trial isolated in a microVM. Vercel announced on September 17, 2026 that Harbor now runs evals on Vercel Sandbox. According to the post, users can run Terminal-Bench or more than 300 other datasets, isolate each trial in its own microVM, keep credentials

Vercel announced on September 17, 2026 that Harbor now runs evals on Vercel Sandbox. According to the post, users can run Terminal-Bench or more than 300 other datasets, isolate each trial in its own microVM, keep credentials out of the sandbox and swap models with a single AI Gateway key.

The announcement treats agent evaluation as an isolation problem as much as a scoring problem. Running each trial in its own microVM keeps state from one attempt out of the next, while holding credentials outside the sandbox limits what an agent under test can reach. A single AI Gateway key for model swaps reduces the setup needed to compare several models on the same dataset.

This matters now because terminal-style benchmarks let agents execute arbitrary commands, which makes shared or credential-bearing environments a real risk during evaluation. Teams running Terminal-Bench style evaluations can hand isolation to managed microVMs instead of maintaining their own execution environments.

The post does not give pricing, concurrency limits, trial startup times or any benchmark results produced on the new setup. Published eval runs on Vercel Sandbox with reported costs and run times would show whether the setup is practical at scale.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Evals](https://newruntime.com/topics/evals.md): Explore the evals topic hub.
- topic hub: [Sandboxing](https://newruntime.com/topics/sandboxing.md): Explore the sandboxing topic hub.
- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)