# Run many sandboxes at once for tests, evaluations or parallel attempts

> Run many sandboxes at once for tests, evaluations or parallel attempts: 0 companies fully do it, 18 with a limit, 24 carry the tag. Every listing in Code execution sandboxes with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/code-sandboxes.parallel-fanout/
- Markdown: https://www.anchorterminal.com/niches/code-sandboxes.parallel-fanout/index.md (~1,750 tokens)
- Slim: https://www.anchorterminal.com/niches/code-sandboxes.parallel-fanout/index.min.md (~380 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/code-sandboxes.parallel-fanout/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 18 | 24 | 9 | 5 |

## What counts as done

50 sandboxes start from the same snapshot without a 429, each returns its result, and the account's concurrency cap and the route to raise it are documented.

Measured, for our own tests: all 50 start within 60 seconds of the first create call.

How agents ask for it: run the test suite split across 20 machines; try five fixes in parallel and keep the one that passes; run this eval over 1,000 tasks; fork the machine at this point

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Agent 37 Cloud (Agent 37 Inc.)**: POST /v1/instances/{id}/fork copies an instance, but a workspace runs 3 instances before its first top-up, 10 after it and 50 only after $50 of top-ups; past the limit a create returns 409 instance_limit_reached.
- **Amazon Bedrock AgentCore Code Interpreter (Amazon Web Services, Inc.)**: Sessions start from a clean runtime, so there is no shared snapshot and packages install on every session; concurrency and API rates are account quotas raised through Service Quotas, with ThrottlingException on 429.
- **Blaxel Sandboxes (Blaxel, Inc.)**: Snapshots capture filesystem, processes and memory and fork into new sandboxes, but tiers start at 10 concurrent sandboxes and rise with 30-day top-up volume, with QUOTA_EXCEEDED past the cap.
- **Celesto (formerly SmolVM) (Celesto AI, Inc.)**: Snapshots and unique-name concurrent computers exist, but no fork-to-N call or concurrency cap is documented for local or Cloud use.
- **Cloudflare Sandbox SDK (Cloudflare, Inc.)**: Each sandbox is a Durable Object you name, and several can start from one snapshot, but no concurrency cap for sandboxes was found on the pages read; the 0.x limits page covers subrequests only.
- **Daytona (Daytona Platforms Inc.)**: Fork is Linux VM only; containers go via snapshots. Tier 1 has 10 vCPU and 10 GiB in total, so 50 sandboxes need Tier 2 (card plus $25). Issue 5137 reports sandboxes unreachable under concurrent verifier load.
- **E2B (FoundryLabs, Inc.)**: Fork snapshots a running sandbox once and boots N copies with memory in a single call, but Hobby allows 10 concurrent sandboxes at 1 create a second; 50 in 60 seconds needs Pro (100 concurrent, 5 a second).
- **Freestyle (Freestyle Cloud Inc.)**: A snapshot captures exact memory and disk and boots any number of clones, but concurrent VMs are 10 on Free, 40 on Hobby and 400 on Pro ($500); one user found 50 restrictive.
- **InstaVM (Checksum Labs, Inc.)**: vms.clone() clones a running VM for fan-out and snapshots boot in milliseconds, but the Free plan allows 5 concurrent VMs and no higher cap is published short of Pro.
- **Maritime (Maritime AI, Inc.)**: Swarms run thousands of sandboxes from one command but need the Scale plan ($500); otherwise machines awake at once are 2 on trial and 5 on Starter, with 429 awake_limit_reached.
- **Modal Sandboxes (Modal Labs, Inc.)**: A memory snapshot restores into any number of clones with processes running, and Offload cut a test run from about 726 to 120 seconds, but the concurrency cap was not on the pages read and a card is required.
- **Morph Cloud (Morph Labs, Inc.)**: POST /api/instance/{id}/branch saves a snapshot and creates `count` children with memory, but no concurrency cap is published; 503s carry Retry-After.
- **SPACE (Perplexity AI)**: Forks and warm pools exist and snapshots are incubating, but the concurrency cap is whatever the operator's cluster and admission limits allow.
- **Runloop Devboxes (Runloop AI, Inc.)**: Disk snapshots create new devboxes, but the trial allows 3 running devboxes and the service ends on 13 November 2026.
- **Sandbox0 (Lougao, LLC)**: Fork and claim-from-snapshot fan out a rootfs, but the default quota is 20 running sandboxes and 5 claims a second, answered with 429 quota_exceeded and Retry-After.
- **Sprites (Fly.io, Inc.)**: Checkpoints snapshot the filesystem and restore the same Sprite; no fork, and creation is limited to 10 a minute on pay-as-you-go, so 50 in 60 seconds needs a plan.
- **Together Code Sandbox (Together Computer, Inc.)**: Any number of sandboxes start from one snapshot, including memory snapshots, but there is no live fork and no concurrency cap is published.
- **Vercel Sandbox (Vercel Inc.)**: Sandbox.fork starts copies from a filesystem snapshot and Pro allows 10,000 concurrent sandboxes at 150 vCPUs a minute, but Hobby allows 10 and 20 vCPUs a minute.

- not this job: DeltaBox Lite (DeltaBox (delta-box)): Concurrent private sessions for several local agents exist, but there is no fleet of sandboxes or fork-to-N.
- not this job: Deno Sandbox (Deno Land Inc.): 5 concurrent sandboxes per organisation during pre-release, raised only by emailing deploy@deno.com; 50 at once is not possible by default.
- not this job: Microsoft Execution Containers (Microsoft Corporation): No fleet, snapshot or fork; runs are local processes.
- not this job: Offscreen (Mark Phelps (individual)): At most two VM jobs at once, by Apple's macOS virtualisation licence.

## What stops agents finishing

- concurrency caps around 50 on entry tiers with a sales step to raise them
- a snapshot build is mandatory before any sandbox (Together)
- Free plans start with zero compute (Morph) or need a card (Modal)

## Around it

Before: code-sandboxes.coding-agent-workspace, agent-harnesses.plan-task. After: decision-models.pick-best-candidate, agent-harnesses.open-pull-request.

## Where this comes from

Weight 9 of 100 in Code execution sandboxes: 5 of 53 mentions: Offload on Modal, a Freestyle thread comment on the 50 concurrent VM limit, a Daytona issue on concurrent verifier workloads, a team testing on three providers, and one builder who evaluated three vendors for per-thread sandboxes.

As data: https://www.anchorterminal.com/api/v1/use-cases/code-sandboxes.parallel-fanout.json, or the MCP tool get_use_case with id code-sandboxes.parallel-fanout. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
