People building autonomous agents do a very ordinary thing. They rent a whole server. They install Python. They leave it running so the packages are still there in the morning.
The agent might work for twenty minutes. The bill is for twenty-four hours. That gap is not a rounding error. It is the product.
The overnight VM
A long-horizon agent is not a web request. It keeps a workspace. It pip-installs a library at 2am. It writes a file it will need on Thursday. If the machine dies, the agent is back to a blank interpreter.
So teams buy a virtual machine “just in case”. The machine is cheap by cloud standards. It is expensive as a night light. Most of the hours you pay for, nothing useful is happening. The Python is merely warm.
Waste is not a mystery
Cloud waste is measured in public. Flexera’s 2025 State of the Cloud report still puts wasted IaaS and PaaS spend at 27%. The same series peaked at 32% four years earlier. FinOps is working. Idle boxes have not vanished.
Flexera’s accompanying 2025 press release is blunt: 84% of organisations name managing cloud spend as their top cloud challenge. Agent builders are not a special species. They just leave more processes running because the process is the product.
What Databricks already tells you
Data platforms learned this the hard way. Databricks’ own cost guidance treats idle interactive compute as a leak, not a lifestyle.
The lakehouse cost-optimisation docs tell you to configure auto-termination on every interactive cluster: after a specified idle time, the compute shuts down. For bursty BI they argue that people leave non-serverless warehouses up because startup takes minutes — and that serverless, which starts in seconds, is how you get both availability and idle termination.
The cost-management blog makes the same cultural point: the expensive habit is a cluster that stays up because someone might come back.
Databricks’ Lakebase guide is even clearer on bursty desks: combining autoscaling with scale-to-zero “can reduce monthly compute costs by 70% or more” for development environments and internal dashboards used only during business hours.
That sentence is from Beyond Provisioning: The Developer’s Guide to Databricks Lakebase Autoscaling. It is about databases, not agents. The economics are the same: if the workload is bursty, paying for a full day of compute is a choice.
On their serverless notebooks, jobs and pipelines, Databricks later claimed efficiency work that produces a greater than 25% reduction in serverless compute cost for most customers, especially short workloads. The number is theirs. The lesson is ours: idle is the first place the bill hides.
A worked example
Take a small always-on box. Amazon lists a Linux t3.medium in US East (N. Virginia) at US$0.0418 per hour. Over a 730-hour month that is US$30.51 if you never stop it — before disk, before traffic, before the afternoon you forget it exists.
The chart below splits that US$30.51 into “useful” and “idle” at three utilisation levels. The heights do not change. Only the colour does. That is the whole trick of an always-on VM: you pay the same whether the agent is thinking or napping.
Figure 1. Illustrative split of a month of Linux t3.medium On-Demand spend (US$0.0418/hr × 730 hours = US$30.51) into useful versus idle time. Utilisation levels are assumptions, labelled as such. Price from AWS T3 pricing, US East (N. Virginia), retrieved for this article. Storage, egress, and CPU-credit overages are omitted.
At 20% useful time you are buying four hours of empty room for every hour of work. Agent teams often sit in that band: a planner that wakes on a cron, a researcher that runs a tool loop, a watcher that mostly waits. The VM is not wrong. The billing shape is.
What to do instead
If the agent is bursty, turn the machine off. Databricks has been saying this in public for years. Pause, auto-terminate, scale to zero. Per-second sandboxes exist for a reason.
If the agent must stay up — days, not minutes; a workspace that should not be rebuilt; children that should be able to exec Python at 4am — then stop pretending a leftover EC2 box is a product. Buy a known hourly rate. Keep the desk. Do not staff a pager for it.
That is the narrow case PySmith is for. We are not claiming your cloud bill will fall 27%, or 70%, or 25%. Those figures belong to Flexera and Databricks, about their worlds. They are here because they name the leak: idle compute that nobody meant to keep.