# Sandbox Selection Brief

Copy this into your coding assistant before choosing infrastructure.

## My workload
- Task the agent must complete:
- Runtime/language and dependencies:
- Expected jobs per day and peak simultaneous jobs:
- Typical and maximum job duration:
- Required CPU, RAM, and disk:
- Files/state that must survive between runs:
- Data sensitivity and allowed network destinations:
- Required regions:
- Monthly budget:

## Instructions for the assistant
Compare two suitable sandbox providers against this workload. Use current official documentation and include direct URLs and the date checked. Separate verified facts, assumptions, calculations, and unknowns. Do not substitute preemptible capacity for on-demand capacity without labeling the tradeoff. Do not treat a one-time credit as a recurring discount or a paid-tier minimum as an additional fee if it covers usage. Flag conflicting documentation.

Estimate the full monthly cost including plan fees/minimums, compute, idle time, retained storage, transfer, and concurrency upgrades. Show formulas and state excluded charges. If a price cannot be verified, label it unknown rather than inventing a number.

Propose a small test using non-sensitive sample data: a successful task, a failed task, and a stop/resume test. Measure completion rate, elapsed time, billed usage, state retention, and cleanup behavior. Rank by cost per successful task and integration fit, not only hourly rate.

Do not create paid resources, enter credentials, upload private data, change production systems, or make purchases without my explicit authorization. Finish with a recommendation, its main tradeoff, and the one unresolved question most likely to change it.

## Evaluation record
- Provider / plan / region / capacity mode:
- Documentation checked on:
- Test task and expected outcome:
- Completed successfully:
- Wall-clock runtime / billed runtime:
- Compute / storage / network cost:
- State retained after stop/resume:
- Cleanup verified:
- Restrictions or unexpected failures:
- Recommendation and remaining uncertainty: