Methodology: how we evaluate agent sandboxes
We list a sandbox if it lets developers programmatically run AI-generated code in an isolated environment. For each, we record documented facts — isolation technology, startup claims, persistence, GPUs, networking, pricing, SDKs, open-source status — from the provider’s own documentation, and we date every entry. Rankings are editorial judgments per use case, based on those documented capabilities, not on paid placement or our own benchmarks.
Inclusion criteria
A product is included if it (1) provides an isolated execution environment, (2) can be created and controlled programmatically, and (3) is marketed or documented for running AI-generated code or agent workloads. Open-source projects and hyperscaler services are included alongside startups.
How facts are gathered
- Primary sources only for product facts: official docs, pricing pages, changelogs and repositories. Each provider page lists the URLs used.
- Performance figures such as cold-start times are vendor claims and are labelled as such.
- When we cannot verify something, we say “not documented” rather than guessing.
- Every entry carries a “verified” date. The whole reference was last reviewed on 2026-11-08.
Evaluation criteria
| Criterion | What we look at |
|---|---|
| Isolation | Documented boundary: microVM, gVisor, container, isolate; multi-tenant suitability. |
| Startup | Documented cold-start and snapshot-restore behaviour. |
| State | Persistence, snapshots, pause/resume, forking, volumes, max session length. |
| Workload range | Languages, custom images, GPUs, desktops for computer use, preview URLs. |
| Control | Network egress controls, secrets handling, resource limits. |
| Deployment | Hosted, BYOC, self-hostable, open-source licence. |
| Developer experience | SDK languages, CLI, MCP server, documentation quality. |
| Pricing transparency | Public per-unit rates, free tier, billing granularity. |
How rankings work
There is no single best sandbox for every workload, so we rank per use case. For each use case on the best-for pages we weight the criteria above by what that workload needs — for example persistence and session length for coding agents, concurrency and forking for RL, GPU access for ML workloads — and order providers by how well their documented capabilities fit. The overall list on the home page reflects breadth of fit across use cases, maturity and adoption signals. Each ranked entry explains why it is placed where it is.
Limitations
- We have not yet run independent benchmarks; performance claims come from vendors. Independent cold-start and reliability benchmarks are on our roadmap.
- Pricing changes often. Always confirm on the provider’s pricing page.
- Fit depends on your workload; use our rankings as a shortlist, then pilot.
Updates and corrections
We review the full reference at least monthly and update individual entries when providers ship notable changes. Providers can submit corrections with a source link; we verify against public documentation before changing an entry.
Cite this page
Agent Sandboxes. (2026). Methodology: how we evaluate agent sandboxes. agentsandboxes.org. Last reviewed November 8, 2026. https://agentsandboxes.org/methodology/
Found an error or a missing sandbox? Submit a correction — every entry links to the sources it relies on.