One person with AI agents can build what used to take a department
We have worked with companies in health, telecom and hospitality. Each one needed a department to get its first product out: engineers, a designer, someone on support, someone chasing invoices. The last product we shipped needed one engineer, a laptop and a handful of agents. That is what the past year looked like.
What changed
Three things changed.
Agents can hold a whole task. An agent can take “add the invoice-matching feature, with tests, and open a pull request” as one instruction. It reads the codebase, makes a plan, runs the tests, fixes what fails. A person reviews the result. That review is the job now.
Tooling caught up. An agent with access to a terminal, a browser and your repository can do most of what a junior engineer does, and it does it at two in the morning without resentment. The gap between “the model can write code” and “the model can ship a change” closed when the tools arrived.
The boring work was most of the work. Writing the privacy policy, generating the sharing images, keeping the sitemap right, answering the same support email. A department existed to absorb that. Agents absorb it now.
What still needs a person
Deciding what to build. Talking to customers. Noticing that the thing the agent built correctly is the wrong thing. Taking responsibility when it breaks.
In practice, the person also needs to be able to read everything the agents produce. If you cannot judge the code, you cannot run a studio on agents; you can only run a slot machine. That is why every S12V Labs project is run by an engineer who reads the code, and why every client talks to that engineer and not to an account manager.
How this shapes the studio
Scopes are small: one workflow, one feature, one product. Agents are fast at the thing you can specify and useless at the thing you have not decided.
Evaluation comes before the build. With a model in the loop, “does it work” is a number from a test set. We build the test set first.
Everything lives in your repository. The output of the work is code, prompts, evaluation sets and runbooks that someone else can pick up. Agents make this cheap enough that there is no reason to skip it.
The person who decides is on the call. The person deciding what to build is the person building it.
What a department still does better
A department of ten still beats one person with agents at some things: parallel customer conversations, on-site work, the long tail of enterprise procurement. For getting a real product in front of paying customers, the first version of almost anything, one person with agents is now cheaper and faster. The studio is built on that.
If you have a product in mind and have been quoted a team and a quarter, tell us about it. The answer is often a person, a few weeks and a working thing.