Launch HN: Agent37 (YC F26) – Low-cost agent sandboxes for running agents

Hacker News Discussion ↗

Low-cost “agent sandbox” platforms like Agent37 are emerging to host AI coding and automation agents, positioning themselves as cheaper, lighter-weight alternatives to traditional cloud VMs. Commenters probe whether this infrastructure is becoming a commodity in a crowded market, questioning the sustainability of competing primarily on price against hyperscalers while also comparing features, cold-start performance, and isolation models (gVisor vs full VMs). A recurring theme is that the harder and more valuable problem may be secure access to user data and services—through scoped credentials, egress controls, and broker layers—rather than the raw sandboxing itself.

Market landscape & competition

  • Many commenters note a crowded “agent sandbox” space, with multiple YC startups (e.g., Maritime, Boat, others) plus incumbents like Fly Sprites, Modal, Daytona, Cloudflare, etc.
  • Debate on whether sandboxes are already commoditized: some argue it’s “just VMs with agent-friendly features,” others say differentiation may come later via specialization.
  • Some see overlap as normal in accelerators when many founders chase the same AI/agent opportunity.

Pricing strategy & sustainability

  • Agent37 emphasizes being significantly cheaper than multiple competitors (claims of ~3x vs Boat, ~5x vs certain AWS instances, ~30x vs Cloudflare, ~94% vs Fly Sprites).
  • Underlying strategy: avoid hyperscalers (rent bare metal from cheaper providers) and exploit spiky agent workloads via denser pod placement.
  • Skeptics worry that competing primarily on price invites a race to the bottom; hyperscalers can cross-subsidize and undercut later.
  • Supporters suggest “carve market share, then get acquired” as a valid outcome. Founder claims current margins are healthy even at low prices.

Technical architecture & performance

  • Uses gVisor-based sandboxes for memory sharing and density; full VMs and “performance” tiers exist for heavier use cases.
  • Filesystems persist across sessions; cold starts are faster from NVMe when warm, slower when restoring from object storage.
  • Auto-sleep and wake are used to keep costs down; free tier limits awake hours per week.
  • Cron is handled via an external cron API that wakes sandboxes rather than in-container cron.

Security, isolation & capabilities

  • gVisor compared unfavorably to full VMs for isolation by some, but many use cases don’t need more.
  • Several commenters argue host isolation is the “easy” part; the hard problems are:
    • Scoped access to user data (Google, email, etc.).
    • Egress allowlists the agent cannot change.
    • Credential brokers that inject real secrets outside the sandbox and enforce per-route/per-action limits.
  • Some explicitly say they’d pay more for strong credential brokering than for cheaper compute.

Developer experience & use cases

  • Early users report easy setup, attractive pricing, but notice cold-start latency on cheapest tier.
  • Popular use case: hosting agent harnesses (Hermes, Claude Code, OpenClaw, etc.) via a unified gateway API that mimics OpenAI’s schema.
  • Questions raised about support for cron-like recurring tasks, bursty agent traffic, filesystem hygiene, and network policy controls.

Misc

  • Interest in BAA signing, SOC 2 in progress, potential open-sourcing, and more payment options (e.g., PayPal).