Agent37, a Y Combinator F26 startup, offers sandboxes for running AI agents at under 10% of typical provider costs. Customers pay an average of $0.37 per month per sandbox, and the company now provides a free tier used by more than 90% of new signups. Infrastructure costs for those free sandboxes average between $0.05 and $0.10 per user each month.

The company began as an OpenClaw wrapper in February and has since added more harnesses. The founder, Vishnu, previously worked as a Senior Software Engineer at Amazon and an Engineering Manager at Quantcast, where he built the internal AI chatbot Diya and the coding harness Ziya. The free plan became possible after a series of cost-optimization techniques were applied to the infrastructure.

How agent workloads shape cost

Agent harnesses follow recognizable patterns when they use CPU and RAM. They need CPU in bursts during initialization or when performing certain tasks, but usage often drops off once a task is complete. RAM demand is higher for activities like browser-based or computer-use work, and usage typically decreases after the task finishes. Efficiently managing these utilization patterns is a key lever for lowering costs.

Another major cost lever comes from where compute capacity is sourced. The founder noted that hyperscalers like Amazon charge 80 to 90% margins on virtual machines, so the obvious alternative is to go one layer down to infrastructure providers such as Hetzner and OVH. Shifting to those vendors alone probably had the biggest impact on reducing expenses.

To make better use of each host, the platform uses gVisor. This technology enables faster boot times and lets the system share available host memory across sandboxes without reserving a full virtual machine's memory for each one. The tradeoff is that gVisor-based sandboxes do not support workloads like Docker-in-Docker or Android emulators.

Because RAM is usually the bottleneck for agent harness workloads—not CPU—the server purchasing strategy focuses on packing more RAM into each host. The company also offers full VM sandboxes for workloads that require them, including the ability to run Docker at a slightly higher price, though these remain low cost given the other techniques applied.

Who's behind the project

Vishnu, the founder, has experience building large-scale internal tools at Amazon and Quantcast. His background includes developing Amazon's AI chatbot and coding harness, which informs the design of Agent37's sandbox platform. The team is actively looking for developers building messaging agents to test their workloads and explore the limits and economics of the platform.

What this means for teams

For groups that need to run AI agents at scale, the cost structure changes the math of what's feasible. A sandbox that costs a fraction of traditional options opens the door to more experimentation and broader adoption within teams. The techniques used—resource pattern awareness, avoiding hyperscaler margins, gVisor utilization, and RAM-optimized host sizing—are specific to agent workloads and may not apply broadly to other compute tasks.

The free plan in particular lowers the barrier for developers who want to prototype agent behavior without committing to a per-sandbox fee. Because more than 90% of new signups are on the free tier, the platform has significant data on how agents actually use CPU and RAM, which feeds back into further optimizations.

Techniques at a glance

  • CPU and RAM usage follows burst-then-idle patterns for most agent tasks
  • Skip hyperscalers; use vendors like Hetzner and OVH instead
  • gVisor enables memory sharing across sandboxes; tradeoff is no Docker-in-Docker or Android emulator support
  • Optimize host purchasing for RAM density, not CPU
  • Full VM sandboxes available for workloads that need them, including Docker support
// Example of how agent sandbox costs compare
// Traditional provider: ~$3.70/month per sandbox
// Agent37: $0.37/month average (free tier: $0.05–$0.10)

The founder's experience building internal AI tools at a major cloud provider gives the company a practical understanding of where costs hide in sandbox infrastructure. For developers evaluating sandbox options, Agent37 offers a way to run agents without the per-sandbox charges that typically limit experimentation.

The platform remains open to developers who want to push the boundaries of what's possible with agent sandboxing and see how the economic model holds up under real workloads.