Sentry founder hunts for sandbox infra to power multi-agent state sharing

Sentry CEO David Cramer (X handle @zeeg) and Cursor co-founder thdxr discussed the challenges of sandbox infrastructure for multi-agent architectures on X. Cramer wants "one sandbox per agent," but running subagents quickly runs into state-sync issues: the parent agent needs to read data produced by child agents, and when child processes run in separate sandboxes, reading data back is "incredibly complicated, expensive, and full of bad words." His concrete requirements: run N sandboxes with independent disks simultaneously, where disks can be mounted to at least 32 other sandboxes at the same time, mount points don't need to be known before sandbox startup, and can change dynamically over the lifecycle — essentially a "1 parent sandbox + N child sandboxes" architecture.

Confirmed

• thdxr noted that read-write disk sharing across sandboxes is one of the most expensive capabilities in public clouds, showing that current cloud provider pricing doesn't match this need.
• Cramer tested Vercel Sandboxes and found they support at most 4 mounts, which can't handle multi-agent architectures.
• Cramer added that multi-host read-write mounts are just the cheapest mechanism he could think of — NFS could also be jury-rigged; the core need is really a shared filesystem from child sandboxes to the parent sandbox. It's not technically hard to solve, but architecturally, if you don't plan for it from the start, it can't simply be patched in later. He also said he doesn't want to put all subagents in one sandbox for convenience, since that would introduce too much system coupling.

Why it matters

• As multi-agent systems spread, "one sandbox per agent" is becoming the mainstream approach for isolation and security, but state sharing and data read-back between parent/child agents has become a new infrastructure bottleneck.
• Existing sandbox products (such as Vercel) fall short on mount counts and dynamic mounting, while multi-sandbox shared read-write disks on public clouds are extremely expensive — exposing a gap on the supply side of agent infrastructure, and potentially a new opportunity for sandbox/cloud providers.

2026-10-17 ~ 2026-10-17 · 9 related posts

• [source] Dev asks for sandbox provider supporting N concurrent sandboxes with dynamically mountable disks — zeeg · 2026-10-17
• One sandbox per agent: why parent-child state sharing is hard — zeeg · 2026-10-17
• [source] Sentry founder: Vercel's 4-mount sandbox limit breaks per-agent sandboxing — zeeg · 2026-10-17
• [source] Cursor founder: RW multi-attach volumes for agent sandboxes are the priciest thing on public clouds — thdxr · 2026-10-17
• Sandbox author: parent-child shared filesystem can be hacked together with NFS, painfully — zeeg · 2026-10-17
• Sentry CEO: reading data back from sandboxed agent child processes is 'complicated and expensive' — zeeg · 2026-10-17
• thdxr challenges zeeg: should parent agents get raw filesystem access to subagents? — thdxr · 2026-10-17
• zeeg on multi-agent design: subagents should expose read-only filesystems to parents — zeeg · 2026-10-17
• Sentry CEO on agent orchestration: let child agents expose their filesystem to parents — zeeg · 2026-10-17