Agent37: Agent Sandboxes at 37 Cents a Month, Because Harnesses Idle
Agent37 did its Launch HN on Monday and the pitch is one number: customers pay an average of 0.37 dollars per month per sandbox, under a tenth of what most sandbox vendors charge, with a refund-the-difference guarantee if you find cheaper. The company started in February as one of the OpenClaw wrapper shops, the founder says so and invites the judgment, then added harness after harness. Over 90 percent of new signups now take the free plan and host their personal Hermes, OpenClaw or Claude Code there, at an infrastructure cost to Agent37 of five to ten cents per user per month. The thread reached 71 points.
The post is really a cost anatomy of agent workloads, which is why it is worth reading even if you never buy. Harnesses burst CPU at init and during certain tool calls, then sit near idle. They need RAM for browser and computer-use tasks, then release it. So the product packs sandboxes on gVisor, sharing host memory instead of reserving a VM's worth per tenant, rents hosts from Hetzner and OVH instead of hyperscalers, the founder cites 80 to 90 percent margins on VMs from time at Amazon, and buys boxes shaped for RAM rather than cores because RAM is the bottleneck for agent workloads. Tradeoff: gVisor sandboxes cannot do Docker-in-Docker or Android emulators, so those go to pricier full VMs. Pricing is per minute of uptime, a 2 vCPU 4 GB box left running all month is 4.76 dollars, a stopped sandbox costs 36 cents for its disk, and egress is firewalled to the public internet only.
The founder previously built Amazon's internal chatbot and its internal coding harness. The ask at the bottom is for people building messaging agents to stress the economics.
This matters because the sandbox is where the weekend's containment debate gets implemented. Nadella wants the brake outside the model, Anthropic cut live internet from evals, OpenAI's grader tried to break its own environment. A per-customer sandbox with a hard disk quota, no host or private-network access, and a spend cap per sandbox is that brake as a line item. Whoever makes the isolated box cheap enough to be the default is deciding how agents get contained in practice.
Launch: https://news.ycombinator.com/item?id=50059096
Site: https://agent37.com
← Back to all articles
The post is really a cost anatomy of agent workloads, which is why it is worth reading even if you never buy. Harnesses burst CPU at init and during certain tool calls, then sit near idle. They need RAM for browser and computer-use tasks, then release it. So the product packs sandboxes on gVisor, sharing host memory instead of reserving a VM's worth per tenant, rents hosts from Hetzner and OVH instead of hyperscalers, the founder cites 80 to 90 percent margins on VMs from time at Amazon, and buys boxes shaped for RAM rather than cores because RAM is the bottleneck for agent workloads. Tradeoff: gVisor sandboxes cannot do Docker-in-Docker or Android emulators, so those go to pricier full VMs. Pricing is per minute of uptime, a 2 vCPU 4 GB box left running all month is 4.76 dollars, a stopped sandbox costs 36 cents for its disk, and egress is firewalled to the public internet only.
The founder previously built Amazon's internal chatbot and its internal coding harness. The ask at the bottom is for people building messaging agents to stress the economics.
This matters because the sandbox is where the weekend's containment debate gets implemented. Nadella wants the brake outside the model, Anthropic cut live internet from evals, OpenAI's grader tried to break its own environment. A per-customer sandbox with a hard disk quota, no host or private-network access, and a spend cap per sandbox is that brake as a line item. Whoever makes the isolated box cheap enough to be the default is deciding how agents get contained in practice.
Launch: https://news.ycombinator.com/item?id=50059096
Site: https://agent37.com
Comments