Agents build fast. A change that used to take a team a sprint can come back in an afternoon. So where does the time go now? It goes to two steps that an agent does not settle on its own: deciding exactly what should be built, and checking that what came back does what the business needs. Both take rounds. Both need someone who knows the business.
The first wait: the specification
A model builds what you describe. If the description is vague or wrong, you get the wrong thing sooner. That makes the description the expensive part.
Take the word customer. Sales may mean a signed contract. The warehouse means a delivery address. Finance means whoever gets the invoice. Nobody notices until the three lists sit side by side in a new tool and nothing matches. A model cannot settle that. People have to agree on one meaning and write it down. We call that the glossary. Add a description of the process, step by step, with each step marked as plain code or as a judgement, and you have a specification an agent can build from.
The second wait: the check
The check is harder than it looks, and Dex Horthy of HumanLayer explains why in Why Software Factories Fail, a talk at AI Engineer. His team ran a software factory with the lights off from July 2025 (8:18 in the video). A few months later they hit a fault no agent could fix, in code nobody had read for three months.
His explanation is about training. Coding models are rewarded when the new test passes and the old tests still pass. Nothing in that reward punishes a design that makes the next change harder: "there's no way in this system that we can penalize it for poor program design" (12:52). So the model learns to make the test green, and the cost shows up later: "the cost function of bad architecture is measured in months and years" (13:23). Another model as judge helps only so far: "if the model knew what good code looks like, it would probably write it in the first place" (14:35).
His conclusion for now is that someone reads the code. Stefan reached the same place from the business side in Someone still has to read the code on spinout.se.
Our addition is about what a green test means. If the agent that wrote the code also wrote the test, green says the code does what the agent understood. It does not say the customer gets what they came for. That needs a check made from the other side.
Why you set the pace
Both steps need the people who know your business: what a customer is, what counts as done, what a good result looks like. So we leave them with you. The agents do not make you wait for us, and no queue of ours sits between your idea and your first version. How fast new features ship depends on how quickly you can say what you want and judge what comes back.
That is also why the pace is yours to raise. Short, clear specifications and quick answers to the checks give shorter rounds. Vague specifications give long ones, whoever builds.
Where we help
Clear specifications, written with our agents. We draft with you: the words, the steps, which steps are code and which need judgement, and what done looks like. Your owner reads it and says yes. Ten clear lines are often enough to start. The point is that the text is exact before anything is built.
Testing, also with our agents. Our agents write automated checks from the glossary and the process description. If the rule says a discount may not exceed 15%, the check tries 20% and expects a refusal. That is the floor. Above it we test what matters most: that the end customer gets the experience they came for, or that the business result is reached. For example, that a customer can place an order and get a confirmation, that the totals on an invoice match the ledger, or that a wholesale buyer sees their own price list. A person on your side judges those results, because only you know what a good one looks like.
What stays with a person
Horthy's advice is to move the human effort earlier: "30 minutes over here in pre-planning and alignment can save you hours in review" (16:50). We agree. One question helps you sort what to leave to agents: when does the cost of a mistake show up? If it shows up at once and a check can see it, agents can run with little supervision. If it shows up months later, as it does in code, a person owns the judgement. Someone on your side, or we on your behalf, should be able to explain the code and change it when it matters.
The short version is on our page How it works. If you want to see what a first specification looks like for one of your processes, book an introduction.
Sources: Dex Horthy, HumanLayer, Why Software Factories Fail, AI Engineer. Quotes and timestamps are from the video's automatic English captions.