AI agents:
what they're for,
how they work,
what they cost.

— The starting point

An AI agent, what exactly is it?

You've seen everyone talking about AI agents for two years. In practice, the word is misused nine times out of ten. I'll give you the definition every serious specialist uses, and which comes from developer Simon Willison: "An agent is a language model — Claude, ChatGPT, Gemini — that runs tools in a loop to reach a goal."

Three words matter: tools, loop, goal.

• Without tools (sending an email, searching the web, writing to a database), all you have is an LLM talking into the void.
• Without a loop (chaining several actions), all you have is a chatbot answering one question at a time.
• Without a clear goal, all you have is a doer with no purpose, who ends up wandering off in all directions.

Anthropic, the maker of Claude, adds a useful distinction (in their reference article "Building effective agents," December 2024):

• A workflow is when you decide the steps in advance. "When a lead lands in Shopify, add it to HubSpot, send it an email, ping me on Slack." It's rigid, it's predictable, it's Zapier or Make.
• An agent is when the LLM decides for itself, at each step, what to do next, based on what it observes. More autonomy, and more surprises too.

— The clear table

Chatbot, workflow, agent: the real differences.

The word "agent" has become a marketing catch-all. Here's the clear table to know what you're really talking about.

	LLM alone	Chatbot	AI agent	Auto workflow
What it does	Answers text	Chats about a specific topic	Acts and chains actions	Chains predefined steps
External tools	No	Sometimes	Yes (email, web, APIs · interfaces that let one program talk to another)	Yes, but fixed
Who decides the steps	The user	Written rules	The model itself	You, in advance
Memory	Current session	Varies	Short and long term	The run's state
Example	"Summarize this text"	Shopify support bot	Claude Code reading your emails and drafting replies	"New lead → CRM → email"

Most tools sold as "AI agents" in 2026 are really beefed-up workflows with an LLM in the middle. Nothing wrong with that. But knowing what you actually have, you spend smarter.

— The anatomy

The four building blocks that make an agent.

A real AI agent is four ingredients put together.

1. The brain

A language model (Claude, GPT-5, Gemini) that reasons, decides, writes. It's what replaces the "if / then" of classic workflows with contextual judgment.

2. The hands

What specialists call the "tools" or "tool use" (plainly: the model's ability to call external functions). Functions the agent can ask to run: sendEmail, searchWeb, readCRM, addCalendar. The agent doesn't perform the action itself. It says "I need to send this email," your code runs the action, and sends the result back to it.

3. Memory

Short term (the history of the current conversation) and long term (a database that stores memories, preferences, the documents it has consulted). Without memory, the agent forgets everything every session, like an amnesiac intern.

4. The decision loop

The cycle that keeps the whole thing running until the goal is reached. It's what lets the agent try, fail, and try again.

My 3 agents in production · measured ROI (April 2026)

So you don't only get third-party numbers, here's what my 3 personal agents save me, measured over 30 days. First-party data, verifiable on my Anthropic bill and my own timesheets.

Total: €38/month for ~ 60 h saved over 30 days. Cost/value ratio ~ €0.63/hour recovered. That's the single most important metric when you're deciding whether an agent is worth it.

— The cycle

How it works, concretely.

An agent's cycle happens in four beats. Always the same.

1. Perceive

The agent reads your request ("organize my Wednesday") and the current state (calendar, unread emails, open tasks).

2. Reason

The model chooses: do I answer directly, or do I need a tool? To organize a Wednesday, it needs to read the calendar and the emails. It decides to call readCalendar() and readEmails().

3. Act

It calls the tools. Your code runs them, sends back the results. The agent sees "Wednesday 10am free, 11am client meeting, 2pm free, three urgent emails, two follow-ups to do."

4. Check

It takes in this information, decides whether the task is done. It isn't yet: it still has to propose a schedule. It drafts the proposal, comes back to you, or creates it directly if you've given it the right to.

The cycle starts again until the goal is reached.

— The family

The four main types of agents.

Not all agents are alike. For an entrepreneur, four categories are enough to understand what's out there.

Type	What it does	Concrete business example
Conversational	Extended dialogue, understands context, uses one or two tools	A lead-qualification bot for a coach — qualifies and books a meeting
Autonomous	Chains several steps on its own for a complex task	A restaurant owner who has it read the last 50 Google reviews, write and post the replies
Orchestrator	Supervises several specialized agents	An e-commerce seller with a "conductor" that delegates to a support agent, a cart-recovery agent, a product SEO agent
Specialized	Narrow expert in one domain	A consultant with an agent dedicated to competitive monitoring across five LinkedIn accounts

An orchestrator agent is the level above. You run into it often in impressive demos. In real life, start with a specialized agent that solves one specific problem.

— The real cases

What it's for, with real examples.

The promise "get yourself assisted by an agent" stays vague until you've seen a concrete case. Here are five publicly documented 2026 examples, with their numbers. I put the source of each figure in brackets so you can cross-check.

Case 1 — A phone line for a restaurant

[vendor Hostie, reported figures]
A restaurant that misses 15 to 20% of calls during the rush, especially between 6pm and 8pm. A voice agent takes the booking, speaks 140 languages if needed, calls the reservation API. Announced result: +141% of covers booked by phone.

Case 2 — Email summaries for a B2B consultant

[internal study GrowthSpree]
Before the agent: 10 minutes per account to read LinkedIn, the news, the website, and write a personalized message. With a Claude workflow: 90 seconds per account. Reply rate 2 to 3 times higher than generic cold outreach.

Case 3 — Abandonment chat for an e-commerce seller

[Snow Teeth Whitening with Rep AI / Shopify]
A chat that re-engages abandoned carts and answers product questions. Over 60 days: 33.85% conversion on abandonment chats, +$220,000 in additional revenue, 50% fewer support tickets.

Case 4 — An orchestrator for a SaaS startup

[Ankor, yachting software, Lindy testimonial]
A small team short on hands. It deploys 20 to 30 Lindy agents. The founder's testimonial: "It was like hiring 10 employees in a week."

Case 5 — A phone line for a tradesperson

[vendor figures GOODEV, Nerolia]
Plumbers, electricians, heating engineers. 30 to 40% of missed calls while on the job = lost quotes. The agent sorts urgency vs. quotes, sends a mobile alert if there's a leak, schedules the slots. Announced gain: 40 to 55% more prospects captured, for €80 to €350 a month in subscription.

— The toolbox

How to talk to it: the system prompt.

The single most important thing to understand when you build an agent is that you don't "talk" to it the way you talk to ChatGPT. You first give it permanent instructions — the system prompt — before any user asks it anything. These instructions define its personality, its limits, its tools.

A structure that works:

ROLE     : You are [a precise persona], expert in [domain].
CONTEXT  : [company, product, typical user].
GOAL     : [expected measurable result].
TOOLS    : You have access to [list of tools + when to use them].
LIMITS   :
  - You never [forbidden action 1].
  - If you're less than 90% sure, ask for confirmation.
  - You never reveal [sensitive info].
OUTPUT FORMAT : [list, email, table…].

A copy-paste example for a personal assistant:

You are Leo, Jérémy's assistant, an entrepreneur running 3 projects.
Your tone: warm, direct, never corporate.
Goal: save Jérémy one hour a day on admin.
Tools: Gmail (read and draft, never send without
approval), Google Calendar (create, edit), Notion (add notes).
Limits: you never delete an event without asking,
you always propose 3 slots for a meeting, you're on a first-name basis.

This system-prompt method is 80% of the work. An agent with a good system prompt and five basic tools will outperform a poorly framed agent with fifty tools. Start simple.

— The paths

The three ways to build one.

You have three possible paths in 2026, depending on your technical level and your budget.

Path	For whom	Setup time	Cost/month	Control
No-code (Lindy, Zapier Agents, Make, n8n Cloud, Gumloop)	You want an agent running tomorrow, no coding	30 min to 2h	€30 to €200	Low to medium
Claude Code + Agent SDK	You can open a terminal and edit a text file	1 to 4h	€5 to €50 (Anthropic pay-as-you-go)	High
Full Python dev (LangGraph, CrewAI)	You code, or you have a dev nearby	2 to 10 days	€20 to €500	Total

The rule of thumb I give the people close to me: start with no-code to test your idea. Switch to Claude Code the moment you get stuck on the logic or your costs go over €80 a month.

— The numbers

What it really costs.

AI API prices in 2026, per million tokens (input and output):

Model	Input	Output
Claude Haiku 4.5	$1	$5
Claude Sonnet 4.6	$3	$15
Claude Opus 4.7	$5	$25
GPT-5	$0.625	$5
Gemini 2.5 Flash	~$0.30	$2.50
Mistral Medium 3	$1	$3

Translated into real scenarios:

Solo entrepreneur — 1 light agent

An email assistant that runs 500 times a month. Stack: Claude Haiku via API, Make.com Core plan at €9 a month, free Vercel hosting. Monthly cost: €15 to €25. Worth it as soon as the agent saves you an hour a month.

SMB — 5 active agents

Support, lead qualification, meeting summaries, inbox triage, quote generation. 50,000 requests a month. Stack: Claude API, a Sonnet and Haiku mix, Supabase Pro $25, Railway $20, Lindy orchestration $50. Monthly cost: €450 to €700. Worth it as soon as it replaces 0.3 of a full-time hire.

10-person team — critical agents in production

Three agents in production, with a searchable document base (RAG), monitoring, regular evaluations. Stack: Claude API with Opus for the complex cases, Pinecone for memory, Datadog for monitoring, multi-region hosting. Monthly cost: €3,000 to €13,000. And you have to add 50 to 100% of the base in hidden costs (integrations, training, security, GDPR assessments).

— The real answer

Does an agent learn from its mistakes?

It's the question everyone asks and that nobody answers honestly. The short answer: no, an agent doesn't learn on its own. The models are frozen after their training. Claude Sonnet 4.6 is the same today as it was a month ago, whatever you told it in between.

But there are four ways you can improve it, from the outside.

1. Adjust the system prompt

You notice the agent keeps slipping up on a type of request. You reword its permanent instructions, you add an example, you spell out a limit. It's free, it's instant, it's always the first thing to do.

2. Give it a memory

You connect the agent to a file or a database where it stores notes between sessions. "Jérémy doesn't like bullet points in his emails." Next time, it re-reads, it applies it. It's what Anthropic calls "Skills" and what most platforms call "memory."

3. Give it documents to consult (RAG)

You build a small library of documents (your product sheets, your sales pitch, your technical docs). The agent consults this library before answering. It's the most effective way to make it accurate about your business. The technical acronym is RAG — Retrieval Augmented Generation. You can ignore it and just remember: a library the agent consults.

4. Fine-tune the model

You retrain the model (plainly: you put the LLM back through a mini-school with your own examples) on thousands of examples of what you want it to produce. It's expensive (several thousand euros and several weeks), it requires quality data, and it's rarely necessary. Reserve it for the cases where the first three levers genuinely aren't enough.

In order, use levers one through three. Only reach for four as a last resort.

— The guardrails

The five risks to know about.

Before you deploy an agent in your business, take two minutes to read these five points.

1. Hallucinations

An agent can "hallucinate" (make up a false answer while sounding sure of itself), cite a source that doesn't exist, get things wrong with total confidence. The 2026 stats put it at 8% hallucinations on average, 18% in customer support, 35% in long multi-turn conversations. Reducible with RAG and human checks, not eliminable.

2. Drift

An autonomous agent can stray from its initial scope (the set of tasks you assigned it), process data it shouldn't see, answer off-topic. Put explicit guardrails in your system prompt.

3. Runaway costs

An agent on Claude Opus with poorly bounded loops can burn through €50 in a day. Always set a cap on calls per user or per day.

4. Security

"Prompt injections" (plainly: booby-trapped instructions slipped into a text the agent will read) are a known attack: someone slips a hidden instruction into an email that your agent runs without knowing. Never give an agent access to irreversible actions (deleting, paying, sending) without human validation.

5. GDPR and the EU AI Act

If your agent processes personal data, you have to run an impact assessment (DPIA · Data Protection Impact Assessment, the GDPR analysis required for high-risk processing), sign an agreement with your provider (DPA · Data Processing Agreement, the standard GDPR contract between you and the processor), minimize the data, and disclose to users that they're talking to an AI (article 50 of the EU AI Act). Fines go up to €15 million or 3% of worldwide revenue.

— The questions that keep coming up

The ten questions everyone asks.

1. What is an agent, in one sentence?

A piece of software that uses an AI model (Claude, ChatGPT) to decide for itself which actions to chain together to reach a goal.

2. Do you need to know how to code to build one?

No, not to get started (Lindy, Zapier, Make offer no-code). Yes, for robust custom builds.

3. Minimum budget to test one?

€50 a month for a first simple agent, €450 to €700 for a serious SMB setup with five active agents.

4. Does an agent learn just by chatting?

No. The models are frozen. For it to improve, you have to give it new tools, new instructions, or an external memory.

5. Cloud or local?

Cloud = fast, pay-as-you-go costs, vendor dependence. Local = private, free after you buy the hardware (around €1,500 of GPU), slower and less capable in 2026.

6. Can hallucinations be fixed?

Reducible with RAG and human checks, not eliminable. Accept that your agent will get things wrong, and put safety nets in place.

7. How long does it take to deploy a first agent?

A day with Lindy or Zapier, two to four weeks in custom code, two to six months for critical production.

8. Is it GDPR-compliant in Europe?

Yes, provided you run an impact assessment, sign a DPA with the provider, minimize the data, and disclose the AI to your users.

9. Which model should you choose in 2026?

Claude Sonnet 4.6 for reasoning, GPT-5 for value for money, Gemini Flash for volume, Mistral for European sovereignty.

10. What's the real return on investment?

An agent that saves you 5 hours a month at €50 an hour is €250 of value. It's worth it as soon as it costs less than that.

— Over to you

And you, are you going to build one?

If you're reading this article with a specific idea in mind — "I could really use an agent to sort my emails," "I'd love a phone line that answers at night" — you're already further along than 95% of the people talking about it.

Three questions to ask yourself before you start:

1. Which repetitive task takes you more than 30 minutes a day?
2. Could an intern do it with clear instructions and two or three tools?
3. Are you ready to give it 10 hours of setup and €50 a month to test?

If you answer yes to all three, you've got a good candidate. Start with no-code, switch to Claude Code once you've validated the use case.

Go further · 3 tutorials that round out this guide

1. Build a Hermes Agent step by step — a dispatch agent that routes incoming requests to the right tool. The "brain + loop + tools" pattern put into practice end-to-end.
2. My Gmail agent that sorts my inbox — the archetypal solo use case: 30 min/day saved, documented stack, GDPR addressed.
3. The agent that reviews my contracts — a concrete "custom Skill" example with memory + RAG. Perfect for understanding how an agent keeps context between sessions.

I can be wrong. If you've deployed an agent recently and your experience is different, write to me. I read everything, I reply. Field feedback is worth a thousand articles.