There is no platform to learn and no prompt to write. You describe the job once. Everything after that is the product: the spec, the build, the connections, and keeping it working as things change.
Ten minutes on the brief form. Name the agent, describe its one job, tick the tools it needs, and say where a human should stay in the loop. Plain English throughout.
Your description comes back as a spec: what the agent will do, what it will not, and a fixed price. Then a short call on the edge cases, because the exceptions are what make an agent useful or useless.
The agent is built and wired into your accounts. Nothing gets copied out of your systems that does not need to be. First runs happen in draft mode so you can read its work before it touches anything real.
You watch the first cycles and tell us where it is wrong. Agents earn autonomy the same way people do: one correct week at a time. You decide when to widen what it does on its own.
Payers change rules. Tools change APIs. Your process changes because you got better at it. Keeping the agent correct is part of the product, not a surprise invoice.
Most teams start with one job and add a second once they trust the first. Each new agent is another description, and the second is faster because your stack is already connected.
Write it the way you would brief someone on their first Monday: here is what lands on your desk, here is what you do with it, here is who to ask, here is what you never do without checking.
A narrow first version is usually two to four weeks. Anything that touches a system without a clean integration takes longer, and we say so during scoping rather than discovering it later.
Price depends on the systems involved, so it is set on a call once the spec is agreed. You get a fixed number in writing before anything is built.
The agent works against your accounts with the narrowest access that does the job. We do not build a shadow copy of your database, and we do not train anything on your data.
That is common in dental, construction, and insurance software. There is usually a path, whether that is a file drop, a portal session, or a workaround the vendor did not advertise. Tell us the system and we will tell you honestly whether it is workable.
No. You need someone who knows the job well enough to say when the agent got it wrong. That is usually the person doing the work today.
You keep your accounts, your data, and the output. We disconnect and hand over documentation for what was built.
The repetitive one that gets skipped when the week gets busy. That is almost always the right first agent.