Agent37 Launches Low-Cost Sandboxes for Running AI Agents
Agent37 launched sandboxes for AI agents at under 10% of most providers' cost, with average customer spending of $0.37 per sandbox per month, according to its Hacker News post.
In the Hacker News post, Vishnu said Agent37 began in February as an OpenClaw wrapper company and later added more agent harnesses. He wrote that offering a free plan was an early goal, but it was too expensive at first because the company was bootstrapped and servers were running on his savings.
The free plan is now available because of cost optimizations, according to the post. More than 90% of new signups are free-plan users who host personal Hermes, OpenClaw, Claude Code and other workloads on Agent37's cloud. The infrastructure cost for those free sandboxes averages about $0.05 to $0.10 per user per month, Vishnu wrote.
Agent37 attributes the low cost to several techniques. Vishnu wrote that agent harnesses consume CPU and RAM in bursts: they need CPU when initializing or performing certain tasks, need more RAM for browser and computer-use tasks, and then usage falls after the task is done. Managing CPU and RAM utilization efficiently therefore lowers cost.
The company also avoids hyperscalers when renting hosts. Vishnu, who said he previously worked at Amazon, wrote that hyperscalers have 80% to 90% margins on VMs, so Agent37 goes one layer down to vendors such as Hetzner and OVH. He called that decision one of the biggest cost levers.
Agent37 uses gVisor to cache and share resources across a host. According to the post, gVisor provides better utilization and fast boot times, and lets Agent37 use available host memory across sandboxes without reserving a full VM's memory for each one. The tradeoff is that gVisor sandboxes do not support workloads such as Docker-in-Docker or Android emulators. Agent37 also hosts full VM sandboxes for workloads that need them, including running Docker, at a slightly higher price.
Vishnu wrote that RAM is typically the cost bottleneck rather than CPU for agent harness workloads, so Agent37 optimizes servers for packing more RAM into a given host. His conclusion from the work, he wrote, is that focusing on a niche sandbox use case and optimizing for it can produce a better cost-performance tradeoff.
Vishnu said he was previously a senior software engineer at Amazon and an engineering manager at Quantcast, and built Amazon's internal AI chatbot, Diya, and internal coding harness, Ziya. Agent37 is particularly looking for developers building messaging agents to test their workloads and challenge the platform's limits and economics.