I build software with AI agents. I describe a feature, the agent writes it, and I read, test, and correct what comes back. SailMOB, a browser sailing game with real wind physics, raced head to head against an AI fleet, was built solo this way. The simulation and the written rules came first, and the agents built against them.
The speed is there, and it moves the hard part of a project somewhere else. If your engineers already use agents, or you're deciding who should build your CRM or your first internal tool, it helps to know where it went.
Writing it down comes first
An agent builds what you describe. A vague description gets you plausible code that does the wrong thing, quickly. Before anything gets built, I write down what we're building and how we'll know it worked: the rules, the edge cases, and the result that counts as a pass. That document carries most of the thinking in the project, and it stays with you.
Testing matters more
Code that looks right is cheap now. Each change gets checked against the written rules by tests that run on every change, and by a person using the thing the way a customer would. The tests are also what let the next person, or the next agent, change the code later without breaking it.
What to ask for either way
Own it. Don't rent it. The repository and the accounts should be in your company's name from the first day, whoever writes the code. Ask for the docs your next engineer will need to run and change what was built, and for the decisions in writing, with the reason for each, so a choice made in week two can be revisited in month six.
What it means for your first build
Agents make a CRM on your own database, or a tracker for the renewals sitting in someone's calendar, much faster to put in front of your team. Agreeing what it has to do still takes judgment, because the agent will build whatever the rules say. That's why the work starts with the Ops Audit, which takes one to two weeks and ends with a written fix plan before anything gets built.