Sandbox comparison · Checked October 27, 2026
Agent37 vs Islo
Keep your agent running. Keep compute costs down.
Agent37 offers low-cost gVisor sandboxes and full VMs. Islo offers persistent microVM environments with gateway policies for agent access to external services. Compare the running cost first, then choose the runtime and controls your workload needs.
Standard sandbox
$4.76
2 vCPU, 4 GB RAM and 4 GB disk for 730 running hours. A low-cost starting point for agents waiting on models and tools.
Full VM for Docker workloads
$11.36
The same explicit 2 vCPU / 4 GB RAM / 4 GB disk size for 730 running hours, with its own kernel. Model and tool charges are separate.
Metered by the minute, without a sandbox subscription. See Agent37 rates.
Agent37 vs Islo: what does running a sandbox cost?
These USD estimates use 2 vCPU, 4 GB RAM and an explicitly selected 4 GB disk. They assume 60 or 730 aggregate running hours, with deletion afterward and no retained snapshots. Credits, taxes, model use and paid tools are excluded.
| Runtime | 60 hours | 730 hours | Runtime tradeoff |
|---|---|---|---|
| Agent37 standard gVisor | $0.39 | $4.76 | Shared host resources; no Docker daemon inside the sandbox. |
| Agent37 full VM | $0.93 | $11.36 | Own Linux kernel; supports Docker and Compose. |
| Islo Machine | $18.17 | $221.04 | MicroVM environment; CPU, memory and disk charged while running. |
Agent37 standard compute: 2 × $0.80 + 4 × $0.70 = $4.40 per 730 hours. VM compute is 2.5 times that. Four GB of disk adds $0.36 per 730 hours. For 60 hours, multiply the total by 60/730.Agent37 billing reference.
Islo lists $0.07 per vCPU-hour, $0.04 per GB-hour of RAM and $0.0007 per GB-hour of disk. At this size: (2 × $0.07 + 4 × $0.04 + 4 × $0.0007) × running hours. Its pricing page says there is no charge when stopped and no seat license fee.Islo pricing.
This is a list-rate comparison, not a throughput benchmark. Equal vCPU and RAM labels do not establish equal performance or isolation. Size disk for your dependencies and data; 4 GB is a pricing example, not a recommendation for every stack.
Keeping files between jobs changes the budget
Agent37: pay for disk while you keep it
Stop or sleep an instance to pause compute charges. Retaining its 4 GB disk for all 730 hours costs $0.36, including time asleep. With 60 running hours during that month, the total is $0.72 for standard or $1.26 for a full VM.
An awake instance still bills compute when CPU use is idle. Enable auto-sleep if your workload can pause, and keep durable files on the billed disk.Review instance lifecycle controls.
Islo: configure pause and cleanup policies
Islo documents idle pause, a total active-runtime ceiling, deletion after a set time, and optional resume on incoming activity. Pausing snapshots state and releases compute while preserving the filesystem for resume.
Lifecycle policies are set at creation and cannot be changed on an existing sandbox. Decide how long a workspace should live before provisioning it. The running-cost table above excludes separately retained snapshots; confirm retention terms for your setup.Islo lifecycle reference.
Choose the runtime and access controls your agent needs
Agent37 for a lower infrastructure bill
Start with standard gVisor when your agent needs shell execution and persistent files on shared compute. Choose a full VM when it needs Docker, Compose or its own Linux kernel. Standard sandboxes cannot run a Docker daemon.
VM creation requires a funded workspace. The Ubuntu VM template defaults to 12 GB disk when resources are omitted, costing $12.08 for 730 hours at 2 vCPU / 4 GB RAM. Its entrypoint has no systemd init; keep container data on the billed disk. The $11.36 example above explicitly uses 4 GB disk.Read the VM requirements.
Islo for gateway policy and VPC requirements
Islo Machines use microVMs with their own kernels and can run services alongside agent code. Islo also offers deployment inside your own VPC; confirm the enterprise terms for that deployment.Islo Machines.
Its gateway profiles match outbound requests by host, with optional path and method rules. They can allow or deny requests and inject credentials outside the sandbox process. If that policy layer is part of your application, evaluate it alongside compute cost before migrating.Islo gateway documentation.
Try one workload before moving the fleet
Pick a real job: clone a repository, install dependencies, run the agent and execute its tests. Measure elapsed time and peak memory. Stop and resume the environment, verify your files, then compare the bill with model and tool spend counted separately. Test required network policies as part of the same trial.
Agent37 is a strong fit when sandbox running cost drives your decision and its runtime supports your workload. Keep Islo on the shortlist when its gateway controls or deployment into your VPC are requirements you would otherwise build.