Which AI agent sandboxes and code execution vendors AI assistants name, Oct 2026
How ChatGPT, Perplexity and Google AI Mode answer 3 buyer questions on AI agent sandboxes and code execution: 15 answers, 19–21 Oct 2026, India.
What this page is. AEOLedger sells answer engine optimization services, so we have an interest in these results. 15 answers (3 questions × 5 runs), collected 21 October 2026 from one location, read by hand. Vendors are named only when an assistant named them; being listed here is a count, not an endorsement or a ranking. The data is downloadable as CSV.
Summary
- Modal (12 of 15), E2B (11 of 15) and one other sandbox vendor (10+ of 15) led the category; E2B was named first in all 5 answers to the general sandbox question.
- The browser-agent question produced a separate set: Browserbase and Browser Use Cloud in 5 of 5, Browserless and Steel in 4 of 5. Perplexity and Google AI Mode also pulled code-sandbox vendors into this question, citing those vendors' own articles.
- Northflank was named in 4 of 6 Perplexity and Google AI Mode answers and in none of 9 ChatGPT answers. In each Google AI Mode answer that named it, Northflank's own article was a cited source.
- 15 further vendors were named in 1 or 2 of 15 answers each.
- 1 further vendor named in 3 to 9 of 15 answers is not shown (see limits).
The 3 buyer questions
Written in buyer wording, without any vendor’s name. Each question was asked in a fresh session every run (runs per question: ChatGPT 3, Perplexity 1, Google AI Mode 1), so 5 answers per question and 15 in all.
| # | Type | Question |
|---|---|---|
| C5-1 | shortlist | “What are the best code execution sandboxes for AI agents?” |
| C5-2 | requirements | “Which sandboxes let a coding agent run a full Linux environment safely with snapshots and persistent files?” |
| C5-3 | requirements | “Which platforms run browser-use agents in isolated environments at scale?” |
Who gets named
Named in X of N: answers whose text recommends or lists the vendor.
Apps, 21 October 2026, Bengaluru, India. ChatGPT (logged out, fresh chat, 3 runs per question); Perplexity (signed in to one account, fresh thread, 1 run per question); Google AI Mode (signed in, personalisation off, 1 run per question). N = 15 answers per vendor.
| Vendor | Answers | Named | By assistant |
|---|---|---|---|
| Modal | 12 of 15 | ChatGPT 6/9, Perplexity 3/3, Google AI Mode 3/3 | |
| E2B | 11 of 15 | ChatGPT 6/9, Perplexity 2/3, Google AI Mode 3/3 | |
| An agent sandbox vendor | — | 10+ of 15 | split not shown |
| Fly.io Sprites | 6 of 15 | ChatGPT 3/9, Perplexity 2/3, Google AI Mode 1/3 | |
| Vercel Sandbox | 6 of 15 | ChatGPT 3/9, Perplexity 3/3, Google AI Mode 0/3 | |
| Browser Use Cloud | 5 of 15 | ChatGPT 3/9, Perplexity 1/3, Google AI Mode 1/3 | |
| Browserbase | 5 of 15 | ChatGPT 3/9, Perplexity 1/3, Google AI Mode 1/3 | |
| Browserless | 4 of 15 | ChatGPT 3/9, Perplexity 0/3, Google AI Mode 1/3 | |
| Northflank | 4 of 15 | ChatGPT 0/9, Perplexity 2/3, Google AI Mode 2/3 | |
| Steel | 4 of 15 | ChatGPT 3/9, Perplexity 0/3, Google AI Mode 1/3 | |
| Cloudflare Sandbox | 3 of 15 | ChatGPT 2/9, Perplexity 1/3, Google AI Mode 0/3 | |
| Kernel | 3 of 15 | ChatGPT 2/9, Perplexity 0/3, Google AI Mode 1/3 |
What gets cited
Domains the assistants linked as sources, counted once per answer.
| Domain | Type | Cited in |
|---|---|---|
| modal.com | cited site | 10 of 15 |
| e2b.dev | cited site | 8 of 15 |
| github.com | cited site | 5 of 15 |
| browserbase.com | cited site | 5 of 15 |
| browser-use.com | cited site | 5 of 15 |
| vercel.com | cited site | 4 of 15 |
| northflank.com | cited site | 4 of 15 |
| browserless.io | cited site | 4 of 15 |
| sprites.dev | cited site | 3 of 15 |
Method
- Fieldwork 19–21 Oct 2026, Bengaluru, India, consumer apps. 3 buyer questions written without any vendor name. ChatGPT was asked each question 3 times (logged out, fresh chat); Perplexity and Google AI Mode once each. 15 answers, every one read in full.
- A vendor counts as named when the answer recommends or compares it, including a row in a comparison table. Vendors mentioned only as an acquisition note, as 'the big names to avoid', or in follow-up suggestions are not counted. Product lines are counted under their company.
The full protocol is on the methodology page.
Limits
- 15 answers per category, 5 per question: enough to say 'named in every answer' or 'in none', not to order vendors in the middle (95% intervals overlap).
- One location (India), no VPN. Perplexity ran in one signed-in account and Google AI Mode signed in with personalisation off; logged-out results may differ.
- Companies we track are not named on this page. As anonymised vendors they appear as 'an agent sandbox vendor' only when named in 10 or more of 15 answers, shown as '10+ of 15' without a per-assistant split, and are otherwise left out; their sites are left out of the sources list.
Data
Download the data (CSV): one row per vendor and per cited domain, with the count, N, assistant split, date and location. Licensed CC BY 4.0: reuse it with a link to this page.
Want this for your own category, with your brand in it? Request a free snapshot.
How to cite and changes
AEOLedger. “Which AI agent sandboxes and code execution vendors AI assistants name, Oct 2026.” AI Shortlist Index v1.0, 22 October 2026. https://aeoledger.com/research/ai-shortlist/agent-sandboxes/
- v1.0, 22 October 2026: first published.
Spotted an error? Write to hello@aeoledger.com. Corrections are dated here.