// guides
Hiring someone to build your AI agent.
The market is young and demos are easy to fake. These are the questions that tell a builder from a demo-maker.
// on this page · 3 sections
Search for an AI agent developer and you get body shops with stock photos, $22-an-hour rate cards and portfolios you cannot click. The gap between a good pitch and a system that still works in month three is where budgets disappear.
One request removes most of the risk. Ask for a system of theirs running in production, unattended, today, with a URL. A case study PDF or a video does not count. A developer selling automation who has never automated their own work is telling you something.
The questions that separate builders from demo-makers
"What happens when it fails?"
A real answer names retries, alerts, fallbacks and a human gate.
The weak answer is that it won't fail.
"How do you test prompt changes?"
Listen for evals, which are automated tests of the AI's output. Saved test cases should catch anything that breaks, and prompts should be under version control.
If prompts live in a dashboard nobody reviews, the agent's behavior is one edit away from a surprise.
"What will this cost to run each month?"
A builder quotes API, hosting and maintenance without hesitating.
A demo-maker has not run anything long enough to know.
"Who owns the code?"
You should, in a repo under your own account.
Where we fit, stated plainly
Anchor 163 is a small senior team with a Fortune 100 platform background, and we run self-operating systems on our own infrastructure. A single automation starts at $5,000, and most scoped builds land between $15,000 and $40,000.
If your project needs a ten-person team, we will say so on the first call and point you to a better fit. Either way, run the URL test on us and on everyone else you talk to.
The architect is ours, running on this site: describe your business and read what it draws. That is one way to judge how we think before you talk to anyone.
Common questions
Ask for a URL to a system of theirs running unattended in production today. Then ask what happens when it fails, how prompt changes are tested, what monthly running costs will be, and who owns the code. Weak answers to any of these predict a failed project.
For prototypes, sometimes. For production agents touching real customers and real data, the failure modes are expensive and usually surface after handoff. Total cost including the rebuild typically exceeds a senior builder’s quote.
You should, in a repository under your own account. That includes the prompts, which should be versioned and tested like any other code.
studio@anchor163.com
// copied