Debate: Persistent VMs vs Ephemeral Sandboxes for AI Agents

On October 11, a technical debate unfolded on X over whether AI Agents should run in persistent VMs or ephemeral sandboxes, led by zeeg and Valtown founder David Crawshaw.

Confirmed

• zeeg took a hardline stance: if you're using persistent VMs for AI Agents, you're doing it wrong; ephemeral sandboxes aren't just the future, they're the present, and they outperform a single or a few persistent VMs across the board.
• He also clarified this wasn't a dismissal of persistent disks, making his position more nuanced than the headline claim.
• Crawshaw quoted zeeg's view and offered his own nuanced position, responding to zeeg's question of "why not use persistent disks."
• Crawshaw explained details of his forked-VM approach: forked VMs have negligible overhead, and idle memory gets swapped out, so he can run dozens of VMs simultaneously on a cheap plan; the base VM is kept alive with systemd.

Why it matters

• An agent's execution environment (degree of persistence, isolation approach, cost model) is a key design decision in today's AI engineering infrastructure. The clash between two frontline practitioners, grounded in real products, gives the community a direct reference: one side advocates a full shift to ephemeral sandboxes, while the other demonstrates that a persistent approach based on systemd keep-alive and forked VMs is equally viable on cost.

2026-10-11 ~ 2026-10-11 · 5 related posts

• [source] Zeeg: persistent VMs for agents are wrong, ephemeral sandboxes are the present — zeeg · 2026-10-11
• Sentry CEO: Persistent VMs for AI agents are wrong, ephemeral sandboxes win — zeeg · 2026-10-11
• [source] Hot take: persistent VMs with agents is doing it wrong, ephemeral sandboxes are the present — davidcrawshaw · 2026-10-11
• Zed and Val Town devs debate persistent VMs vs ephemeral sandboxes for AI agents — zeeg · 2026-10-11
• [source] Val Town founder: branch VM overhead is negligible, base VM kept fresh with systemd — davidcrawshaw · 2026-10-11