Re:Launch HN: Agent37 (YC F26) – Low-cost agent sandboxes fo...
by Ace7-coder on Thursday October 22, 2026 @08:52PM
(#50203164)
This exact scenario—silent failures leaving compute spinning—is what drove me crazy managing sandbox infrastructure.
You think a task failed and terminated, but the EC2 instance or EBS volume is still sitting there accruing charges. I got so tired of hunting these down manually or writing nested boto3 loops that I built an open-source CLI to automate it. It pulls the raw state into Pandas, flags the 0% I/O resources, and auto-generates a bash script so you can just review and delete the waste.
If you are running high-volume jobs like this and want a safety net to catch the orphans, it might save you some time: https://github.com/Ace7-coder/aws-finops-auditor-lite"
Your point about bursty CPU matches what I see: install and build spike for 20–60 seconds, then the process mostly idles while the harness reads output. RAM only matters the day a project pulls in a headless browser.
The things that actually cost me time were not CPU or RAM, though. First, network policy. I run untrusted code with an egress allowlist, and most launched projects assume open internet, so the useful knob is per-sandbox allowlist editing with a proxy for package registries, not more memory. Second, environment hygiene. The sandbox needs a clean env with no inherited secrets, and a persistent sandbox that does not pick up changed env vars or allowlist entries without a restart bit me more than once. Third, a wall-clock timebox per run, because a stuck npm install behind a proxy fails silently rather than loudly.
Two questions: can a sandbox’s egress allowlist be changed while it is running, and when gVisor does not support a workload (you mention Docker-in-Docker), does the API fail fast or does the job just hang?