AI agent development services: agents that run inside your business, not in a demo

What is an AI agent, and how is it different from a chatbot or a workflow?

A workflow is a fixed sequence: when a form arrives, do step A, then B, then C. It is predictable and cheap, and it cannot handle a case nobody drew on the diagram. A chatbot answers questions in a chat window; the good ones read your documents before answering, but they stop at the reply. An agent is different in four ways. It has tools: it can search your CRM, read an email thread, query a database, create a ticket or draft a quote. It has memory: it keeps the context of the case across steps and, when needed, across days. It makes decisions: given the goal and what it found, it chooses the next action instead of following a branch someone hard-coded. And it runs in a loop: act, observe the result, decide again, until the job is done or a stopping rule fires.

That loop is what makes agents useful and what makes them dangerous when built carelessly. A workflow that fails, fails visibly at a step. An agent that is given the wrong tools or no stopping rules can do the wrong thing confidently and repeatedly. Most of the engineering in AI agent development is not the prompt; it is deciding which tools the agent gets, which actions require a person, how the agent knows it is done, and how you see what it did.

Which AI agents actually pay off for a small or mid-sized business?

After enough discovery calls a pattern shows up. The agents that earn their cost do one narrow job that a person today does many times a day, with a clear definition of done and a clear place to escalate. The ones that disappoint are the "assistant for everything" that impresses in a meeting and that nobody trusts with a real customer.

Five types come up most often, and they are the ones I quote with confidence:

How do I build AI agents, and on what stack?

The model is the smallest part of the decision. I build on the Claude API (Anthropic) and the OpenAI API, and use Google Gemini where it fits, chosen per task on quality, cost and latency rather than loyalty. The agent's tools are defined as functions that call your actual APIs: the CRM, the inbox through the Gmail or Google Workspace APIs, WhatsApp through Meta's Cloud API, your ERP, your database. The agent never gets a generic "browse the internet" tool when a precise "look up customer by email" tool will do; narrow tools are safer and cheaper.

The runtime is either n8n, self-hosted or cloud, when the agent lives naturally among other workflows and your team wants to see the flow, or a Python service with FastAPI when the logic is heavier, needs tests, or has to handle volume. Memory and retrieval live in Postgres with pgvector, on Supabase or your own instance, so the agent can remember the case and search your documents with the same database you already back up. It runs on Railway, Vercel, GCP or AWS in your accounts, containerized, with secrets managed properly.

Two things are non-negotiable in every build. First, human approval before anything irreversible: sending to a customer, invoicing, deleting, changing a record that other systems depend on. Second, evals and logging: a fixed set of test cases the agent must pass before any prompt change goes live, and a log of every run with the inputs, the tools called and the output, so you can audit what it did and I can improve it with evidence instead of guesses. I run Claude-based agents in my own operation every day on exactly this setup.

What does an AI agent build look like, step by step?

Take the most common request: an intake agent for a company where every inquiry that arrives through the web form, a shared inbox and WhatsApp is read by the owner or by one salesperson. Here is what the build looks like in plain words.

What do AI agent development services cost?

I do not bill hourly for builds. After a free discovery call you get a written proposal with a closed scope and a single fixed price, delivered in three weeks for a scoped agent and three to five weeks for larger ones. The reference points are the two productized systems published with list prices on this site: the AI Intake Agent with knowledge base at USD 4,900 (optional USD 1,200 per month for operations and iteration), and the AI Operating System at USD 6,800 (optional USD 1,800 per month). A custom agent that is close to one of these is priced the same way; a scope that is much smaller or much larger is quoted as a fixed price after discovery, never as an open-ended hourly engagement.

Infrastructure and API subscriptions (LLM usage, hosting, n8n Cloud if you choose it) are billed to you directly, in your accounts, with a cap declared in the proposal. Because every run is logged, you can see the exact model cost per case instead of guessing. The 60-day operational guarantee covers anything I built that breaks after handoff, at no cost.

Why do AI agents bought from a demo fail in production?

A demo agent is optimized for the meeting: it answers the five questions the seller rehearsed, on the seller's data, with no consequences if it is wrong. Four failure modes show up when the same thing meets real customers and real records.

When should you not build an AI agent?

An agent is the wrong tool more often than sellers admit. If every case follows the same steps, a workflow in n8n does the job for a fraction of the cost and with none of the ambiguity. If the volume is a handful of cases per week, a person with a good template is cheaper than any system. If the data the agent would need is not in any system yet, because it lives in someone's head or in unstructured chats, the first project is getting it into a CRM or a database, not the agent. And if your CRM already has the feature, use it. I do not resell tools, and I will say any of this on the first call.

I also decline scopes where the agent would take irreversible actions with no human in the loop from day one, or where the promise is "replace the team". Those builds fail, and the reputational cost lands on you. What works is an agent that removes the reading, the looking-up and the first draft, and leaves the judgment where it belongs until the logs show it can be trusted with more.

How does the engagement work?

It starts with a free 15-minute call to see whether an agent is the right tool and whether I am the right person. If there is a fit, you receive a written proposal with closed scope, timeline and one fixed price. During the build you get weekly updates and approve before anything touches production or a customer. Handoff includes documentation, the exported code and workflows in your repository, Loom walkthroughs so the system survives staff changes, and the 60-day guarantee. Optional monthly support after that covers iteration, new tools for the agent and monitoring.

I work remotely from Chile with companies in the US, Canada and Europe, in English or Spanish, with full overlap with US business hours. It is a solo practice; for large scopes I bring in a second engineer, and you always work directly with the person who builds.