Search for AI agent development companies and you will find hundreds of agencies that did not exist eighteen months ago. Some are engineering teams that happen to be young. Most are marketing shells around no-code tools, one model API, and a sales deck. This guide gives you a filter: the four provider types, the signals that separate production shops from demo factories, and the exact questions to ask on the first call. We build agents for a living, so read this knowing where we stand, but every test below works against us too.
§ 01The four types of providers
Almost everyone selling agent development falls into one of four buckets, and the bucket predicts your outcome better than any portfolio page.
- Freelancers. Fast and cheap for prototypes. The risk is not skill, it is coverage: one person cannot monitor production around the clock, and continuity ends when they take a better contract.
- No-code agencies. They assemble agents from platform building blocks. Fine for internal, low-stakes automations. The ceiling shows up at the edges: unusual inputs, API changes, error recovery. When the platform breaks, they file a support ticket like everyone else.
- Platform resellers. They sell you a subscription to someone else's product with a setup fee on top. You are not buying development, you are buying configuration. Ask yourself whether you could configure it without them in a week.
- Engineering studios. They design architecture, write code, own infrastructure, and operate what they ship. More expensive than the first two buckets, and the only bucket where anyone is accountable when the agent fails at 2 AM.
None of these are wrong in general. They are wrong or right for a given job. If the agent touches customers, money, or regulated data, you are shopping in the fourth bucket, and the rest of this guide is about telling real studios from agencies dressed as studios.
§ 02Five signals of a production-grade studio
- Production evidence with failure history. Ask for their longest-running agent and, more importantly, what broke during that period and how they handled it. A provider who claims nothing ever broke has never run anything in production. Real operators tell war stories.
- A written architecture process. Before code, there should be a documented design: what the agent decides alone, what escalates to a human, which systems it touches, what happens when those systems are down. We publish our own approach in our engineering notes, and any serious studio can show you an equivalent.
- Confidence handling. The single most revealing technical question: what does the agent do when it is not sure? The only acceptable answer is some version of escalate to a human with full context. An agent that always answers is an agent that confidently makes things up on your customers.
- Infrastructure and data ownership. Insist on a dedicated, isolated, encrypted instance per client. If your agent shares a database or a server with other customers of the agency, one bug is a data breach.
- A concrete operations offer. Deployment is the midpoint of the engagement. The monthly fee should name specifics: monitoring, error alerts, prompt and logic updates, model upgrades, cost reporting. A vague maintenance retainer means nothing is actually being maintained.
§ 03Red flags that end the conversation
- A precise price before a discovery call. Custom scope quoted blind means a template being resold.
- Guaranteed accuracy figures. Nobody who understands these systems guarantees 99 percent anything. Serious providers talk about escalation design instead.
- Demo-only portfolios. Screenshots and videos, no clients you can name or reference calls you can request.
- Agreement with everything you propose. Some processes should not be agents. A provider who never pushes back is selling, not engineering.
- Pressure to sign before scoping. Real studios have discovery pipelines, not month-end quotas.
§ 04How pricing should be structured
Healthy engagements share a shape: a discovery conversation first, a scoped proposal second, a deposit to start the build, and a monthly operations fee once deployed. The build is a project, the operation is a service, and both are priced separately. We covered the full cost breakdown, including what is included in each phase, in our guide to AI agent development services.
One structural note on geography. US agencies quote 25,000 to 100,000 USD for scope that European engineering studios deliver from 1,500 EUR plus monthly operation. This is labor arbitrage, the same reason half of Silicon Valley's code is written in Eastern Europe already. The evaluation criteria above matter far more than the flag on the invoice, and they are exactly how you verify that a lower price is arbitrage rather than corner-cutting.
§ 05The first call, scripted
You need thirty minutes and five questions.
- What is your longest-running production agent, and what broke during that time?
- What does the agent do when it is not confident?
- Who owns the infrastructure, and is my instance isolated?
- What exactly does the monthly fee cover, item by item?
- Have you ever told a client not to build an agent? What did you recommend instead?
Take notes on the fifth answer especially. It is the cheapest integrity test available to a buyer, and it fails more vendors than the other four combined.
──
If you want to run this script against us, we will take the call. Thirty minutes, your process, honest answers, including the answer that you might not need an agent at all. See what we have shipped first if you prefer to check the evidence before booking.