Agent sandboxes / SFCompute Autoresearch Sandboxes alternative

An SFCompute sandbox alternative for returning Linux agents

SFCompute Autoresearch now offers agent and reinforcement-learning sandboxes alongside interactive CPU and GPU nodes. Its sandbox path already snapshots memory, processes, page cache and disk and supports stateful forks. The separate interactive-node path only persists a configured home directory; those two contracts must not be combined.

Start buildingCompare capabilities ↓

Reviewed 2026-10-19 · Published by the Sandbox0 team

When SFCompute Autoresearch Sandboxes fits

SFCompute fits workloads needing stateful microVM forks, saved execution memory and integrated interactive GPU/CPU research nodes. Its agent sandboxes are a direct runtime comparison, with real RAM and disk snapshot capabilities.

When to consider Sandbox0

Consider Sandbox0 when the agent needs a whole durable writable RootFS with enabled automatic online growth and supported files while guest compute is explicitly paused. Compare the required process-memory lifecycle, storage access and billing conditions against SFCompute’s sandbox model.

Capabilities

SFCompute Autoresearch Sandboxes and Sandbox0: what persists?

Feature comparison with documentation and verification status
Environment and runtime lifetimeSFCompute Autoresearch SandboxesNot established in this reviewSandbox0ttl and hard_ttl default to 0 (disabled). Configure idle pause and deletion policies for the same persistent sandbox identity; an idle running sandbox still accrues compute charges.Read documentation →
Persistent filesSFCompute Autoresearch SandboxesAgent sandbox snapshots retain microVM disk; separate interactive CPU/GPU nodes retain only the configured home directory.View evidence ↗Sandbox0The writable RootFS survives successful checkpoints, including software, system configuration, dependencies, and user data.Read documentation →
Whole RootFSSFCompute Autoresearch SandboxesBaked agent sandbox state includes initialized microVM disk and OS state; do not apply this to the home-only node product or separate runtime mounts.View evidence ↗Sandbox0Persistence extends beyond /workspace to paths such as /usr/local, /etc, and /home. Runtime tmpfs and virtual filesystems are excluded.Read documentation →
Release compute, keep stateSFCompute Autoresearch SandboxesStateful sandbox snapshots can restore execution state; automatic idle parking returns spare memory and can leave resident work running and billed.View evidence ↗Sandbox0Pause compute between tasks; resume on available capacity. Default resume starts fresh processes from the saved filesystem.Read documentation →
Memory and process stateSFCompute Autoresearch SandboxesThe sandbox environment explicitly snapshots memory, processes and page cache along with disk, enabling stateful forks.View evidence ↗Sandbox0Experimental memory checkpoints can preserve compatible process state. Use memory: true on both pause and resume on a compatible deployment. A filesystem checkpoint alone does not preserve running processes.Read documentation →

Online RootFS growth

On growth-enabled images, Sandbox0 automatically expands the writable RootFS as free space runs low, within a configured sparse logical ceiling up to 64 TiB. Agent processes keep running. Growth is enabled in the qualified Cloud region; confirm activation for other regions. Existing images and snapshots retain their captured storage geometry. The ceiling is logical capacity, not preallocated host disk.

SFCompute sandbox shapes are fixed at bake and inherited by forks; a whole-disk automatic growth contract is not established. Its documented 64 GiB top shape refers to RAM, not a disk-capacity ceiling. Sandbox0’s enabled online RootFS growth is a separate storage capability.

Files while compute is paused

Sandbox0's Files API reads persistent files without resuming compute. Filesystem-only pauses also permit edits; retained-memory pauses are read-only and reject mutations with 409. SSH, SFTP, file watching, and runtime-only mounts require running compute.

SFCompute’s documented stopped-node write helper wakes compute before the write can be retried; sandbox execution similarly uses a wake lifecycle. Sandbox0 supports direct paused RootFS operations: filesystem-only generations allow mutations, while retained-memory generations are read-only.

See paused file operations →

Stateful sandbox snapshots already include memory and disk

Sandbox environments bake initialized microVM state including memory, running processes, page cache and disk. Forks are copy-on-write, and running sandboxes can be snapshotted without stopping them. Durable snapshot copying proceeds in the background. Interactive nodes differ: only their configured home directory persists, so tools installed elsewhere can reset after stop.

Source 1 ↗Source 2 ↗

Automatic parking returns spare memory rather than freezing all work

Idle sandboxes can park after roughly 1.5 seconds without an in-flight exec. Parking returns spare memory; resident models and background processes can remain active and keep consuming billable memory. Wake and fork check available memory, and a failed wake does not start the requested work. This is not a promise that every parked sandbox has zero memory charge.

Source 1 ↗Source 2 ↗

Shape and billed memory require careful comparison

Sandbox sizes are fixed when baked and inherited by forks. The medium shape has 4 vCPUs and 8 GiB guest RAM but bills 16 GiB, unlike its visible guest-memory number. Larger shapes also share memory with /tmp. Compare the billed GiB and active/parked portions rather than multiplying every shape’s guest RAM by one rate.

Source 1 ↗Source 2 ↗

Interactive nodes have a different idle and file lifecycle

CPU/GPU nodes use their own time meter and idle-stop grace periods, with current prices returned by the API and applicable calendar discounts. A file-write request to a stopped node wakes it without writing and must be sent again after it is running. That behavior must not be described as mutation of a stopped filesystem.

Source 1 ↗Source 2 ↗

Pricing

Compare billing models, then price your workload

CPU allocation, memory, billing minimums, disk, plan fees, and supported workload differ. These meters are not equal-performance machine quotes. Confirm region, resource units, and included allowances before estimating a total.

SFCompute Autoresearch Sandboxes

Pricing reviewed 2026-10-19

Sandbox active memory
$0.04572 per billed GiB-hour
Use billed shape memory, including the medium shape’s 16 GiB billing basis.
Rate source ↗
Sandbox parked state
Approximately $0.000139 per GiB-hour
Resident active memory can still incur the active rate; parking does not freeze every process.
Rate source ↗
Resident snapshot storage
$0.10 per deduplicated GiB-month
Rate source ↗
Cold snapshot storage
$0.05 per GiB-month
Fork fee is $0; baking uses ordinary resource time.
Rate source ↗

Active and parked portions, billed shape memory, snapshot residency and deduplication affect sandbox costs. Interactive CPU/GPU nodes use separate time-based rates with calendar discounts. Compare the rate returned for the selected node and the billed sandbox memory independently.

Check current SFCompute Autoresearch Sandboxes pricing ↗

Sandbox0

$0.015/GiB-hour compute (billed by the second at approximately $0.000004167/GiB-second), plus $0.02/GiB-month of retained incremental RootFS and memory checkpoint data. Shared base images and platform recovery retention are excluded; paused sandboxes accrue storage charges only.

Compute / GiB-hour
$0.015
Retained data / GiB-month
$0.02
Paused compute
$0.00

A 4 GiB agent, 100 running hours

$6.00 compute

4 GiB × 100 hours, billed by the second. Keeping 20 GiB of billable incremental data for a full billing month adds $0.40. CPU is derived from memory by the platform; choose a tier that fits your workload.

Understand storage billing →

Before migrating from SFCompute Autoresearch Sandboxes

  1. Separate agent/RL sandboxes from interactive CPU/GPU nodes, and inventory baked system state, memory, snapshots and home-only node files.
  2. Export required disk state and port compatible Linux execution, retaining GPU nodes and model workflows where necessary.
  3. Test required process continuity, whole-system retention, growth and paused files; compare billed memory and snapshot storage with actual workloads.

Common questions

Is Sandbox0 a drop-in SFCompute Autoresearch Sandboxes replacement?

Evaluate API compatibility and product scope. A shared sandbox category does not make SDKs, lifecycle semantics, or limits interchangeable.

Does keeping files also keep a running agent?

Filesystem state and volatile process state are separate. Sandbox0 requires explicit compatible memory checkpoints to preserve running processes; its default filesystem resume starts fresh processes.

Sources and review scope

The Boat team's Battleships directory is the discovery source. References below describe the product and plan discussed on this page. Check current availability, regional terms and billing units before migrating. Where the cited material does not establish a capability or price, the comparison identifies that uncertainty.

Battleships research record ↗autoresearch.sfcompute.com/docs ↗autoresearch.sfcompute.com/llms.txt ↗autoresearch.sfcompute.com/ ↗Sandbox0 RootFS capacity →

Related comparisons

Agent 37 vs Sandbox0 →AgentComputer vs Sandbox0 →ainclave vs Sandbox0 →Alibaba Cloud Agent Sandbox / FC / AgentRun vs Sandbox0 →Alibaba Cloud AgentBay vs Sandbox0 →Amika vs Sandbox0 →