The Pulse
Agent37 Pitches Persistent Sandboxes as a Cheaper Agent Runtime
Agent37 used an Oct. 13 Hacker News launch post to pitch persistent, isolated sandboxes for AI agents, with usage-based billing and a free tier. Its founder says average customer spending is 37 cents per sandbox each month, while the compan
AI.info Team · Written with AI ·
Agent37 says its customers spend an average of 37 cents per sandbox each month; a Hacker News commenter questions whether an infrastructure startup can build a durable business by selling cheap compute against well-funded cloud rivals. The Oct. 13 launch post puts that disagreement alongside a product pitch: persistent, isolated environments for AI agents, billed by use rather than as fixed monthly seats.
“Our customers spend an average of $0.37/month per sandbox.”
Vishnu, founder of Agent37
That average is not a flat subscription price. Agent37’s published rate estimate for a standard sandbox with 2 vCPUs, 4 GB of RAM and 4 GB of persistent disk is $4.76 for 730 running hours. When the instance sleeps or stops, compute billing ends, but keeping its disk costs $0.36 for 730 hours; model and tool usage are separate.
Files persist; compute stops when the sandbox sleeps
Agent37’s pitch targets teams that give each customer an agent of their own. A sandbox keeps files and connected tools across sessions, sleep and restarts, and can wake when a new request arrives. The company says developers can start from templates for Hermes, OpenClaw, Claude Code and Codex, or bring a custom harness or Docker image.
The service also offers a browser and integrations with apps including Gmail, Slack, Notion and GitHub through Composio. Agent37 says its hosted environments use gVisor for isolation and that the platform provides lifecycle controls, backups and spending limits. Those are company descriptions of the offering, not independent security or performance results.
Low prices come with workload limits
In the Hacker News post, founder Vishnu says the company lowers costs by matching infrastructure to agent workloads, whose CPU demand often rises in bursts, and by renting servers from providers such as Hetzner and OVH rather than relying on hyperscalers. He says gVisor lets Agent37 share available host memory across sandboxes without reserving a full virtual machine’s memory for every instance.
That approach does not fit every task. Vishnu says the gVisor-based sandboxes do not support Docker-in-Docker or Android emulators; Agent37 also offers full virtual machines for workloads that need capabilities such as running Docker, at a higher price. The company’s pricing page cautions that its comparison is a list-rate estimate, not a performance benchmark, and that workload, storage needs and competing providers’ allowances can change which option costs less.
Hacker News users challenge the business case
One commenter argued that large cloud providers could subsidize sandbox features to sell higher-margin services, making it difficult for a smaller company to capture value. Another questioned whether competing mainly on price would hold up if established providers cut rates. Those comments do not establish how Agent37 or its competitors will fare, but they put the startup’s low-cost pitch in direct tension with questions about differentiation and long-term economics.
Vishnu acknowledged that Agent37 faces a crowded market, saying his Y Combinator batch includes more than five companies doing similar work. He suggested that providers may eventually specialize in different areas. The company’s earlier Y Combinator launch page, listed 14 days before the Hacker News post, describes the same product, so the Oct. 13 post is a new pitch to that forum rather than the service’s first public appearance.
The next test is real agent workloads
Agent37 says more than 90% of its new signups use the free plan, and that free sandboxes cost it about 5 to 10 cents per user each month. In a Hacker News reply, Vishnu said the free tier limits awake time to about 10 hours a week; he also said the company restarts memory-heavy sandboxes when a host reaches critical memory pressure and moves new or waking instances to relieve it.
The pricing argument now rests on usage, not just the headline average. Agent37 is asking developers building messaging agents to try their workloads and test the platform’s limits. For free users, the current boundary is about 10 awake hours per week.