Custom AI agents: what they are and when you need one
A chatbot answers. An agent acts: it decides what to do next, calls tools, and finishes a task without a person steering every step. Most businesses need a chatbot for customer conversations and an agent for the work behind them, and the two are built very differently. I build both, on the Claude and Anthropic API with MCP for tool access, and I run them in production.
What is a custom AI agent?
An AI agent is a program that is given a goal rather than a script. It decides what to do next, calls tools to do it, reads the result, and keeps going until the task is finished or it hits a boundary you defined. Custom means it is built around your data, your tools and your rules, instead of a generic product assuming how your business works.
That autonomy is the whole point and also the whole risk. An agent that can book, refund, send or delete is useful exactly because nobody has to click through the steps, and dangerous for the same reason. Good agent engineering is mostly about what the agent is not allowed to do.
How is an agent different from a chatbot?
| AI chatbot | AI agent | |
|---|---|---|
| Given | A message | A goal |
| Does | Answers, maybe retrieves a document | Plans, calls tools, checks results, repeats |
| Ends when | It has replied | The task is done or it escalates |
| Typical use | Support answers, FAQs, guidance | Booking, data work, outreach, back office |
| Failure mode | A wrong answer | A wrong action |
| Needs | Good content and retrieval | Scoped tools, guardrails, logging |
Not sure a website assistant is worth it for your business? The missed inquiries calculator puts a number on it. Most real systems use both. fahrchat looks like a chatbot to the customer, because WhatsApp is the surface, but the part that matters resolves an availability question against a calendar and books a real appointment for a paying driving school.
When should a business build a chatbot instead?
When the work ends with the answer. Product questions, opening hours, policies, onboarding guidance, internal knowledge lookups: all of that is a retrieval problem, not an autonomy problem. A retrieval-augmented chatbot over your own documents is cheaper to build, far easier to evaluate, and it cannot take a wrong action because it has no tools. If someone sells you an agent for that, they are selling complexity.
How does an agent workflow get built?
- Map the process by hand. Whatever a person does today, written down step by step, including the exceptions. This step finds most of the reasons a project fails.
- Define the tools. Each action the agent may take becomes an explicit, narrow tool: read the calendar, create the booking, send the message. Nothing is implicit, and MCP is what makes these tools reusable across systems rather than glue code per project.
- Give it the context. Retrieval-augmented generation for your documents, plus the state the agent needs to know where it is in a conversation or a job.
- Set boundaries. Which actions require confirmation, which never happen automatically, what triggers a handoff to a person, what happens when a tool call fails or the model returns nonsense.
- Test with real inputs. Real messages, real edge cases, adversarial ones included. Playwright where something has to be driven through a browser.
- Ship narrow, then widen. One channel, one use case, full logging, then more once the failure modes are known.
- Operate it. Scheduled runs, retries, alerts, and a view where a human can see what the agent did and why.
Why Claude and MCP?
The reasoning layer here is Claude and the Anthropic API, and tool access runs over the Model Context Protocol. MCP is the part that changes the economics: instead of hand-wiring every integration into every agent, a tool is written once as a server and reused. Claude automation in practice means the model plans and calls those tools, while ordinary TypeScript, Node.js or Python code does everything deterministic, because you should never ask a model to do what a function can do reliably.
Around that: Supabase for data, Vercel for hosting, Twilio and the WhatsApp Business API for messaging, cron jobs for anything unattended.
What agent systems have I actually built?
- fahrchat, a live WhatsApp assistant that books real appointments for a paying driving school, on Twilio and Stripe.
- JobMachine, a fully agentic outreach system that scores prospects, drafts and sends cold outreach and tracks replies, running daily and unattended. Private, not a product.
- Nexus, a personal multi-agent system with orchestration, recursive memory and cross-machine operation. Also private, shown as proof of engineering ability.
- aimentionsyou.com, which queries ChatGPT, Claude, Perplexity and Gemini to track how often a brand is mentioned, and turns that into a report.
- AIRESHAPE, a Blender plugin built for a client that orchestrates Gemini, Kling and Meshy into one video pipeline.
What does custom AI agent development cost?
No packages and no price list. The drivers are how many tools the agent touches, how autonomous it has to be, how bad a wrong action would be, and whether you want it handed over or operated. You are quoted after a short scoping conversation by email.
If the agent is one piece of a wider automation need, start at AI automation services. If you want ongoing capacity rather than one system, that is fractional AI engineering. Otherwise write to hello@marcomori.net or use the contact section, and tell me what the agent would take off your plate.
Questions people ask
What is the difference between an AI agent and a chatbot?
A chatbot responds to a message inside a conversation. An agent has a goal, decides which steps to take, calls tools such as your calendar or database, and stops when the task is done. A chatbot tells a customer the slot is free; an agent books it.
Do I need an agent or a chatbot?
If the value is in the answer, build a chatbot. If the value is in the action that follows the answer, build an agent. Many systems are both: a conversational surface for the customer and agent logic behind it, which is how fahrchat works.
What is a custom AI agent, compared to an off-the-shelf one?
A custom agent is built around how your business actually works: your data, your tools, your rules about what it may never do. Off-the-shelf agents assume a generic process and break at the edges where your business is specific.
Which stack do you build agents on?
Claude and the Anthropic API for reasoning, MCP for tool access, retrieval-augmented generation when the agent needs your documents, and TypeScript, Node.js or Python around it. Supabase, Vercel, Twilio and the WhatsApp Business API cover most of the infrastructure.
What is MCP and why does it matter?
The Model Context Protocol is a standard way to give a model access to tools and data sources. It matters because it turns every integration into a reusable server instead of bespoke glue, so the same tools can be shared across agents.
How do you stop an agent from doing something wrong?
By narrowing what it is allowed to do. Tools are scoped, destructive actions are gated or require confirmation, everything is logged, and anything outside the defined path escalates to a person instead of improvising.
Can an agent run unattended?
Yes, if it is built for it. JobMachine scores prospects, drafts and sends outreach and tracks replies on a daily schedule without supervision. Unattended operation is a design decision, not a setting you switch on afterwards.
What does custom AI agent development cost?
It is quoted after a short scoping conversation by email, with no packages. The drivers are the number of tools the agent touches, how autonomous it has to be, and how much failure handling and logging the use case demands.
Start with an email. Describe the process that eats your time. I reply with a few scoping questions, then a written scope and a fixed quote. No sales call, no packages.