July 25, 2026

•

9 min

ByDiego Carrion·Co-founder, Duotach

AI agentsClaude CodeDefinitions

What is an AI agent

An AI agent is a system that takes a goal and executes the steps to meet it using real tools. A chatbot answers; an agent acts. Here is the full definition, the five components it needs to actually operate, and the cases where you should not use one.

How to build an agent with Claude Code See an agent in production

The definition

An AI agent is a system that takes a goal expressed in natural language, decides on its own which steps to take to meet it, and executes them using real tools: reading and writing files, querying databases, calling APIs, running code. It does not return a recommendation about what to do. It does the thing.

The boundary with a chatbot is exactly that. A chatbot explains how to issue an invoice. An agent issues it, verifies it went through, and tells you if it failed. The language model is the same; what changes is that the agent has hands.

The one-line test

If the system finishes its work by returning text, it is a chatbot. If it finishes having changed the state of something (a file, a database row, a ticket, an invoice), it is an agent.

Agent, chatbot and classic automation

The three get confused constantly and solve different problems.

	Chatbot	Automation (n8n, Zapier)	AI agent
What it produces	Text	One fixed action	A sequence of actions decided at runtime
Who defines the steps	A prior script	A flow drawn in advance	The system itself, per case
Handles variable input	Barely	No	Yes, that is the point
Predictable	Fairly	Completely	Not entirely: hence verification
Running cost	Low	Low	Medium to high
When it fits	Frequently asked questions	Stable, repetitive processes	Processes with constant exceptions

The practical rule

If you can draw the entire process as a flowchart and the chart closes, build it with classic automation: cheaper, faster, and it does not fail creatively. An agent earns its cost when the chart never closes, because a new case always shows up.

The five components of a production agent

If one is missing, what you have is a demo.

1. A model that reasons

The engine that interprets the goal and plans the steps. The most visible part and almost always the cheapest line in the project.

2. Tools wired into real systems

Without access to the ERP, the mailbox, the database or the billing system, the agent can only write about what it was told. This is where MCP comes in.

3. Memory and context

Knowing what happened in previous runs, which documents it already processed, and the rules specific to the company. Without this it repeats work and contradicts earlier decisions.

4. Explicit permissions

What it can read, what it can write, and what requires human approval. An agent without defined limits is an incident waiting for a date.

5. Verification

Something that checks the result before accepting it: tests, validation rules, or a human on the final step. The component most often skipped and the most expensive one to skip.

Of the five, the one that decides whether the project reaches production is almost never the model. It is the second and the fifth: actually connecting the systems, and being able to prove the output is right.

What MCP is and why it changed the cost of agents

MCP (Model Context Protocol) is an open standard created by Anthropic to connect language models with external tools and data sources. It works like an adapter: an MCP server exposes a system (the ERP, the CRM, a database, a repository) and any compatible agent can use it without a custom integration.

Why it matters in money terms: previously each new agent meant re-coding access to each system. With MCP, the ERP is integrated once and serves every agent that comes after. The first automation is still expensive; the fifth is much cheaper. That is the difference between a project and a platform.

What it costs to run an agent

Three separate costs. The classic mistake is budgeting only the first.

Cost	Reference	Comment
Model (license)	Claude Pro 20 USD/mo (17 billed annually) · Max from 100 USD/mo · Team 25 USD/seat/mo (20 annually)	Anthropic pricing as of 2026-07-25, in USD, tax not included. The smallest and most predictable cost.
Implementation	Quoted by scope	The bulk. Driven by how many systems must be integrated and how clean the input data is.
Maintenance	Recurring	APIs change, formats change, and the agent needs adjusting. An unmaintained agent degrades.

Be wary of a fixed price quoted before anyone has looked at your operation. The real range depends on how many systems need integrating and the state of your data, and neither is knowable without looking.

When NOT to use an agent

The limits almost nobody publishes.

The process is already solved by a simple rule

If deterministic automation covers it, adding an agent adds cost and variance with no upside. More intelligence is not better if the problem does not need it.

The error is severe, irreversible and unverifiable

Transfers, data deletion, client-facing messages without review. If there is no automatic way to confirm the action was correct, either a human stays in the loop or it does not ship.

The input data is disorganized

If documents are unnormalized and every department has its own format, the real project is a data project. An agent on dirty data amplifies the mess.

Nobody will own running it

An agent needs an owner who watches the output and corrects course. Without that person it degrades within weeks and ends up switched off.

Agents running today

Three of our systems in production, what they do and what was measured.

Agent system for SEO and GEO

A system of agents that researches, writes, audits and validates content optimized for search engines and language models, with validators that halt the process when output does not pass. See the case.

Proposal generator for a media agency

Builds proposals and media plans in minutes that previously took hours of manual work. See the case.

RAG knowledge base on AWS

A private knowledge base with semantic retrieval, in production for a company in Ecuador. See the case.

In automation projects with published metrics, measured results were 15 hours per week recovered in pharmacy inventory management and an 85% reduction in report preparation time for a marketing agency.

Frequently asked questions about AI agents

What is an AI agent?+

An AI agent is a system that takes a goal in natural language, decides on its own which steps to take, and executes them using real tools: reading and writing files, querying databases, calling APIs, running code. The difference from a chatbot is that a chatbot answers and an agent acts. A chatbot explains how to issue an invoice; an agent issues it.

How is an AI agent different from a chatbot?+

In three ways. A chatbot returns text and stops; an agent executes actions against systems and verifies the outcome. A chatbot follows a conversation script defined in advance; an agent decides the sequence of steps at runtime based on what it finds. And a chatbot fails visibly (it answers badly), while an agent can fail silently (it performs the wrong action), which is why it needs verification and scoped permissions.

Is an AI agent the same as an n8n or Zapier automation?+

No. Classic automation runs a fixed sequence: if A happens, do B then C. It is deterministic and does not adapt. An agent decides the sequence per case, which matters when inputs vary (documents in different formats, requests written any which way, frequent exceptions). The practical rule: if you can draw the whole process as a flowchart and the chart closes, use classic automation, which is cheaper and more predictable. An agent earns its cost when the chart never closes.

What does an AI agent need to work inside a company?+

Five components: a language model that reasons, tools wired into the company real systems (usually via MCP or an API), memory or context so it knows what happened before, explicit permissions defining what it may touch, and verification that checks the result before accepting it. If any of the five is missing, what you have is a demo, not a production agent.

What is MCP and why does it matter for agents?+

MCP (Model Context Protocol) is an open standard created by Anthropic to connect language models with external tools and data sources. It matters because it solves the most expensive problem in agent work: instead of building a custom integration per system, an MCP server exposes that system once and any compatible agent can use it. It is the difference between integrating the ERP once and integrating it for every new agent.

What does it cost to run an AI agent?+

Three separate costs. The model: Claude Pro is 20 USD per month (17 billed annually), Max starts at 100 USD per month, and Team is 25 USD per seat per month (20 billed annually), all in USD with tax not included. Implementation: the bulk of it, driven by how many systems must be integrated and how clean the input data is. Maintenance: systems change, APIs break, and the agent needs adjusting. The most common budgeting mistake is pricing only the first one.

When should you NOT use an AI agent?+

When the process is already solved by a simple, stable rule, because an agent adds cost and variance with no upside. When an error is severe and irreversible and there is no automatic way to verify the result. When the input data is so disorganized that it has to be cleaned first, in which case the real project is a data project. And when nobody in the company will own running it: an agent without an owner degrades within weeks.

Do AI agents replace people on the team?+

In practice they replace tasks, not roles. The processes where they pay off are repetitive and high volume: classifying documents, moving data between systems, handling first-line questions, generating recurring reports. The pattern we see in projects that work is that the person moves from executing to supervising and resolving exceptions, which is where their judgment is worth something.

Do you have a process an agent could run?

The useful conversation does not start with the tool, it starts with the process: which one is costing you the most hours or the most errors today.

How we work

We start with a scoped assessment that defines what to automate first and with which technology, including the option that you do not need an agent at all.

Tell us about your process How to implement Claude Code

Seguir leyendo