• Sources: primary
  • Summary: Akshat Bubna describes sandboxes that run the agent harness separately from the containers where tool calls execute, with shared volumes and snapshots of a root file system or a single directory so state persists for days without paying for idle compute. Ramp's internal coding agent data showed CPU and memory use varying widely within and across sandboxes, so provisioning for peak would mean about 15 cores and 15 GB per sandbox. He says domain allow lists are not enough and users write sidecar proxies with path-level rules and inject credentials the agent cannot read, and that Modal sees 20 million snapshots and 2.6 million volumes created a week.
  • Why it matters: Teams running agents in sandboxes can plan for per-path egress rules, hidden credentials, snapshot-based persistence and burstable resources instead of fixed peak-sized containers.