What is an AI agent?

That ability to decide is what makes an agent useful, and it's also why it needs limits. In a company, an agent writes to customers and changes data on behalf of the business. At Xarxalia we work with agents on that basis: there is always a person who can review what they do and stop them.

Agent or workflow?

In its guide "Building effective agents", published on 19 December 2024, Anthropic separates two kinds of system. In a workflow, the model and the tools follow paths defined in code. In an agent, the model directs its own process and decides which tools to use and when.

The same guide recommends looking for the simplest solution possible and adding complexity only when it's needed. An agent's autonomy costs more and lets errors compound, so it has to be tested in a sandboxed environment and given guardrails.

In practice, plenty of repetitive work in a company fits in a workflow. An agent makes sense when the cases vary too much to write every branch, for example a customer email that could be a complaint, a booking change or a pricing question.

How an agent works, step by step

1. It receives a message or a goal through whatever channel: WhatsApp, email, a form, a phone call.
2. It looks up the context it needs: information about the business, the customer's history and the rules it has to follow.
3. It chooses the next action and runs it with a tool it has permission to use: reply, create a record, suggest a date.
4. It reads the result and decides again. Anthropic stresses that the agent should check at every step what actually happened (the tool's output) instead of assuming it.
5. It stops when it reaches the goal, when it hits a limit or when it needs a person.

One point that often gets overstated: in most production systems the model does not change with use. What improves is what you give it: knowledge of the business, approved answers and rules. If a vendor says their agent "learns on its own", ask what exactly changes and who reviews it.

Why a business agent needs human oversight

The EU's artificial intelligence regulation, Regulation (EU) 2024/1689, covers this in two articles:

• Article 14 requires high-risk systems to be overseeable by people. Those people must be able to decide not to use the system, to override or reverse its output, and to halt it with a stop button or a similar procedure.
• Article 50 requires that anyone interacting with an AI system is told so, unless it is obvious. It applies from 2 August 2026.

Not every agent is high-risk, but the principle behind Article 14 holds just as well for one that serves customers: a person has to be able to see what it did, stop it and undo it.

How we do it at Xarxalia: four levels of autonomy

The MIA digital employees work at four levels of autonomy, set per employee, per channel and per type of action:

1. Suggests. Drafts the reply and leaves it ready. Sends nothing until you approve it.
2. Answers what you've approved. Replies on its own to the cases you've signed off and holds the rest.
3. Answers and logs. Works on its own and records everything so you can review it whenever you like.
4. Acts within limits. Takes and completes actions without going beyond the margin you've set.

At any level, the agent stops and alerts a person if it doesn't know the answer, if it detects a complaint or an emergency, if there's legal or financial risk, if the conversation turns tense, if the customer asks for a human or if the decision isn't its to make. Every action is logged with a date and the person responsible. Every project goes live at the lowest level, and the company decides when to raise it and can lower it at any time. The details are on Technology and trust.

How agents connect: MCP and A2A

An agent can only act in the systems it's connected to. Two open standards cover that connection:

• MCP (Model Context Protocol). Anthropic introduced it on 25 November 2024 as a standard for connecting AI assistants to the systems where data lives: content repositories, business tools and development environments.
• A2A (Agent2Agent). Google introduced it in April 2025 and on 23 June 2025 handed it over to the Linux Foundation. It lets agents from different vendors discover what each one can do and split up tasks.

The HUB, Xarxalia's platform, implements both. At hub.xarxalia.com it publishes its A2A agent card (protocol version 0.3.0) and an MCP server (version 2025-06-18). With the company's key, an external assistant can query its search data, reports, knowledge base, CRM and tasks. Without the key, it can't find out which agents the company has or talk to them.

What an agent does well in a company

Work that repeats every day and follows clear rules. That's what the MIA digital employees cover:

• MIA Reception collects messages from WhatsApp, email, the website and the phone, answers with the business's information, creates a record for each new contact and decides who to pass it to.
• MIA Sales picks up the quotes nobody has followed up and flags real interest.
• MIA Back office sorts email and the calendar and sets aside what a person needs to decide.
• MIA Bookings checks availability, confirms and handles changes.
• MIA Billing issues invoices, chases unpaid ones and flags what's running late.

The aim is for the same team to serve more customers, and serve them better, because they stop spending the day on repetitive work. Anything involving legal or financial risk, or a decision the rules don't cover, always goes back to a person.

LET'S TALK

Frequently asked questions