What an agent actually does

Gartner expects up to 40% of enterprise applications to carry a task-specific agent by the end of 2026, up from less than 5% the year before. The same firm also expects more than 40% of agent projects to be cancelled before the end of 2027, for want of demonstrated value, controlled costs or adequate risk controls.

The two forecasts do not contradict each other. It is the clearest movement in the market, and it is also where the gap is widest between what is announced and what is deployed. The same firm has a name for the part of the noise that is not work: agent washing, the rebranding as an agent of an assistant, a robotic-process bot or a chatbot already in place.

The difference is in the path, not the intelligence

An automated flow follows a rule written in advance: if this, then that. It always does the same thing, and that is its principal quality.

An agent is given an objective, looks at the actual situation and chooses the steps. It is therefore useful where the rules would be too numerous to write down, and unsuited to cases where they must be followed to the letter. A regulatory filing does not need an agent: it needs a flow that never deviates.

A flow

1. A request arrives through the form
2. If the amount is over €500, alert the manager
3. Otherwise, record it in the CRM
4. Send the acknowledgement

The rule is written in advance. The path is always the same, and that is exactly what is asked of it.

An agent

1. Objective: the request gets handled
2. Looks at what the request contains and what it has to hand
3. Picks the next step, runs it, looks at the result
4. Repeats until the objective is met

The steps are chosen at run time. Two similar requests can take two different paths.

That starting point decides almost everything else, and it is settled by looking at the task, not at the tool.

What agents actually in service look like

They are simple and tightly framed. They do one thing, within a narrow scope, with limited rights, and they most often run on the same flow engine as the rest of the company’s automation, n8n or its equivalent.

The architectures discussed at conferences are almost never the ones a company puts into production. Gartner places agents that collaborate with one another in 2027 and agent ecosystems in 2028, and estimates that only about one hundred and thirty vendors, out of the thousands now claiming to be agentic, genuinely are.

We stand on the side of what gets deployed. An agent with access to everything and authority over everything is a demonstration, not a working tool.

Three limits set before a line is written

Rights. An agent never has more access than its task requires. Read everywhere and write nowhere is a good starting point, and you open up precisely what is needed afterwards.

The trace. Every action is logged with what triggered it and what it produced. Without that, explaining an anomaly becomes impossible, and the first incident is enough for the whole thing to be unplugged.

The human hand. Whatever commits the company, a send, a payment, a publication, stays triggered by a person. The agent prepares, the person confirms. That is not excessive caution: it is what allows the scope to widen later, because trust is built on weeks without surprises.

It reads

Wherever it helps the task: the inbox, the CRM, internal documents.

It writes

Only where it has been opened up, one source at a time, and every write is logged.

It commits

Never alone. A send, a payment, a publication are prepared by the agent and triggered by a person.

Where to start

With a task someone already does, whose volume and cost are known, and where an error can be undone. Sorting incoming requests, preparing a file, formatting something before approval.

Starting with the most impressive case is the surest way of putting nothing into service.