An agent that runs for weeks is not a function. It is a worker with a desk. The desk has a Python on it. The Python has packages. The packages took twenty minutes to install and the agent would like them still to be there on Monday.
Most of the market sells either a hotel room that is reset every night, or a building you have to plumb yourself. Neither is a desk.
Two bad desks
Serverless sandboxes are excellent at the thing they were built for: untrusted code, short jobs, a filesystem you can throw away. Pause and resume is a real feature. Per-second billing is honest. Then you come back and the room has been remade. Common libraries are gone. Long-lived child processes are a conversation with the vendor’s session cap.
Raw VMs are the opposite failure. The desk survives. So does the patching, the SSH key, the disk that filled up, the security group you meant to tighten. You wanted Python. You bought a data centre square.
Agents do not care which cloud account the desk sits in. They care that import pandas still works, that a file written on Tuesday is readable on Friday, and that another agent can exec code without someone opening a ticket.
What staying up actually costs
Fixed hourly rates look dear when the process is mostly idle. They look simple when it is not. The chart below is a labelled benchmark, not a bake-off with a trophy.
We price a 2 vCPU / 4 GiB shape — the same size as PySmith Medium — using public unit rates collected on 17 Sep 2026:
- Modal Sandboxes: US$0.00003942 per core-second and US$0.00000667 per GiB-second. Running: about US$0.38/hr.
- E2B usage: US$0.000014 per vCPU-second and US$0.0000045 per GiB-second, on top of plan fees. Running: about US$0.17/hr.
- Northflank allocated compute: US$0.01667 per vCPU-hour and US$0.00833 per GB-hour. Allocated: about US$0.067/hr.
- PySmith Medium (illustrative launch price): US$0.08/hr, billed while the instance exists, busy or not.
Per-second lines assume you only pay while the process runs. The PySmith line is flat because that is the point of a desk: you know the number before the month starts. Plan fees, storage, and GPUs are omitted. E2B Pro still documents a 24-hour continuous sandbox cap; weeks-on is a different product.
Figure 1. Monthly cost (730-hour month) for a 2 vCPU / 4 GiB shape at 10%, 25%, 50% and 100% duty cycle. Modal, E2B and Northflank use published unit rates as of 17 Sep 2026 (Modal, E2B, Northflank). PySmith Medium is an illustrative fixed SKU at US$0.08/hr. If the agent is mostly off, per-second wins. If it must stay up, the fixed line stops moving.
Read the chart in that order. At 10% duty, Modal and E2B are cheaper than a desk you never turn off. Northflank’s allocated rate is already in always-on territory and undercuts our illustrative Medium SKU. We are not the cheapest computer on the internet. We are a Python that stays up, with an MCP socket, at a number you can put in a spreadsheet.
The Databricks habit
Data teams were trained out of “leave it running”. Databricks’ cost-optimisation guide puts auto-termination on every interactive cluster. Idle time is treated as a defect. For databases that only work office hours, their Lakebase write-up says scale-to-zero can cut compute 70% or more.
That habit is correct for bursty notebooks. It is the wrong habit for an agent whose job is to still be there. Turning the desk off every time the human looks away is how you get a clean bill and a stupid agent.
The useful import from Databricks is cultural, not a cluster setting: know which workloads should die when idle, and which should not. Then price them differently.
A desk is a product
PySmith is the second kind. You get a persistent interpreter, a workspace that keeps packages, and an MCP connector so agents exec Python without SSH. The hourly rate does not change because a child process spawned. Isolation is a system container, not a microVM. Burst to a hundred thousand sandboxes is someone else’s chart.
If you need a hotel room for an hour, rent a hotel room. If you need a data centre, you already know the vendors. If you need a desk that is still there next week, that is a smaller shop — and it should cost a number you can say out loud.