In the businesses I run there is no DevOps engineer. Deploys to all our projects and zones are done by AI agents. We kept our former CTO on a small retainer as a safety net (in case an agent breaks something serious). So far we haven’t called him once.
I’m a medical doctor by education and a CEO, not an ops person. So this is not a guide to a toolchain. It’s what the setup looks like from the side of the person who answers for it, and where I still don’t let an agent close the loop alone.
The Agents Work in Shifts, Not in a Chat Window
The first thing that changed was where the agents live. They don’t run on my laptop. They run on a schedule on separate servers, and each business line has its own shift. A shift starts by the clock, does its job and stops. I only start things.
All of it runs on subscriptions (three Claude plans at $200 a month each), not on per-token API billing.
What is on the shifts today — deploys, our own CRM (the agents built it from scratch, and they also work inside it), email sorting per business line, meeting scheduling and technical support.
Why Technical Support Was the Easy Part
Technical support works on agents for a simple reason. Technical questions have checkable answers — a login fails, a page doesn’t load, a setting is wrong. An agent can find that, apply a known fix and then confirm it worked. The confirmation step is the whole point. If there is no way to check, I don’t give it to an agent.
The One Rule for Deploys
Nothing ships because an agent finished. It ships because it passed a check.
This sounds obvious until you see how an agent reports a wrong result. It doesn’t hesitate. A broken change comes back with exactly the same confidence as a good one (there is no “I’m not sure” in the output). So we don’t read the agent’s own summary as evidence. The check is separate from the work, and the result goes further only after it passes.
The second rule is about access. An agent gets access for one task, not a standing role. When the task is over, the access is over too.
Every Workflow Has a Named Human
An agent can do the work, but it can’t be accountable for it. So every agent workflow in my company has a named person who owns the result. That person answers for it exactly as if a colleague had done the work.
So when something goes wrong, the owner of that workflow answers first. The retainer CTO stays behind that for the serious cases (and again, not needed so far).
What Went Wrong
The biggest problem was not the agents. It was my tasks.
A vague task given to a person gets clarified over coffee. Given to an agent, it gets done — confidently and wrong. For ops this is dangerous because the output looks finished. So I write every task with four parts — the goal, the limits, what “done” means and how we will check it. If I can’t write the last part, the task isn’t ready for an agent.
The second lesson is about rework. If a person has to redo the agent’s work every time, the agent saves nothing — the work just moved to another desk. When I see that, I close the scenario and don’t try to rescue it with a longer prompt.
Where the Limit Moved
Before, my limit was people and the job market. At the moment the limit is Claude tokens and computing power (memory and GPUs). That’s why I rent separate servers, including GPU machines, instead of running agents on a laptop that chokes.
And most of the tokens don’t go where people expect. In my own measurement, more than 85% of the modeled spend went toward handling context — reading and caching information — rather than generating new output (DOI 10.5281/zenodo.22759216). So short shifts that do one thing and stop are also an ops decision, not only a cost one.
What I’d Tell a Team That’s Thinking About It
Start with one function where the result is easy to check (technical support was that for us). Name the person who owns it before the first run. And count the cost of a task that was done and accepted, with reruns and review time, not the price of the tool.
I think memory is becoming the new gold for this kind of work.
Evgenii Arsentev, MD, PhD, is CEO of AskDocDoc and founder of arsentev.ai. A medical doctor by education, he runs his businesses without full-time developers, DevOps or designers — the work is done by AI agents on a schedule, with a named person accountable for every result. ORCID 0000-0002-9120-7298.

