Agent 37

Agent 37 is a YC-backed startup whose core offering is not a single agent application but a persistent sandbox cloud environment for running AI agents. Based on publicly available information, the product centers on five capabilities. First, it ships ready to use, with mainstream agents such as Claude Code, Hermes, and OpenClaw pre-installed, so users can skip environment setup. Second, the sandboxes are persistent: unlike ephemeral sandboxes that are destroyed after each task, Agent 37 retains your files, code, and configuration until you explicitly delete them. Third, billing is usage-based — compute is charged only for the minutes an agent is actively running, while stored files incur a small storage fee. Fourth, deployment is simplified to a single command, removing the need for DevOps expertise. Fifth, agents can securely connect to everyday tools including Slack, Gmail, and GitHub.

It is worth noting that the company was founded in 2026 and is at a very early stage. Details about the underlying architecture, sandbox isolation mechanisms, and security model are not publicly documented, so the above description relies largely on official claims rather than independent verification.

The first use case is hosting coding agents such as Claude Code. Developers running agents locally often face constraints around compute and environment management; a persistent cloud sandbox allows tasks to span long sessions with work state recoverable at any time, which is useful for teams running multiple agents in parallel.

The second is workflow automation connected to internal tools. With Slack, GitHub, and Gmail integrations, agents can take on process-driven tasks such as assisting with pull request reviews, summarizing messages, or organizing email — provided the underlying third-party agents are reliable enough for those tasks.

The third is experimentation by individuals and small teams. For users who want to explore AI agents without maintaining servers, the per-minute billing model keeps the cost of trial and error relatively contained, making it a reasonable entry point for early adopters of the agent ecosystem.

Officially, deployment takes a single command and the pre-installed agents eliminate environment configuration, so the initial setup barrier is low. In practice, two sources of friction remain. First, users still need a working understanding of how coding agents operate; tools like Claude Code deliver the most value to users with some engineering background and prompt-writing skill. Second, because the company is so new, community resources, tutorials, and troubleshooting channels are likely sparse, which raises the cost of resolving issues independently. The quality of documentation is currently unknown.

For non-technical users, authorizing and managing app connections to GitHub and similar services also carries a learning curve. On balance, the product is better suited to developers and early-stage teams with technical backgrounds than to complete beginners.

Public pricing information is limited. The company describes its model directionally: compute is billed per minute of active agent runtime, and a small fee applies to files kept in persistent storage. Specific rates, free tiers, or subscription plans have not been disclosed, so readers should check the official website for current details.

The model has clear advantages. Compared with a monthly cloud server, it is cheaper for low-frequency use since idle time costs nothing in compute. Two caveats apply, however. Persistent sandboxes mean storage fees accumulate as files build up, so unused environments should be cleaned up periodically. Additionally, long-running coding tasks can consume a large number of billed minutes, meaning heavy users may not see meaningful savings. Running small test workloads first to measure actual consumption is a sensible approach.

Agent 37 targets the infrastructure layer of the AI agent ecosystem. The combination of persistent sandboxes and per-minute billing addresses two real pain points in today's agent tooling: disposable environments and paying for idle compute. Pre-installed mainstream agents and app integrations further lower the barrier to entry, making the product attractive to developers and early-stage teams.

The risks are equally clear. Founded in 2026, the company is extremely young with limited public information, and its stability, data security practices, and long-term viability remain unproven. The user experience also depends heavily on third-party agent ecosystems such as Claude Code, which Agent 37 does not control. Overall, this is a reasonable option for technically capable early adopters willing to tolerate the uncertainty of a brand-new product. Organizations considering production use should adopt a wait-and-see stance until the product matures and its pricing and security mechanisms become more transparent.