Agents that carry out the work
Getting a good answer does not move the work forward if someone then has to open three pieces of software to apply it. An agent reads the request, decides how to proceed, and does the work inside your tools. With a trace of every action.
The difference with ordinary automation lies in the path. An automated flow follows a rule written in advance: if this, then that. An agent is given an objective, looks at the actual situation and chooses the steps. That is what makes it useful where the rules would be too numerous to write down, and dangerous where they must be followed to the letter.
What ships, and what gets announced
It is the clearest movement in the market, and it is better to look at what is being deployed than at what is being announced. The agents actually in service 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 your automation, n8n or its equivalent. The architectures discussed at conferences are almost never the ones a company puts into production.
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, and it shows at the first incident.
What we put around an agent
- A written scope: what it is allowed to do, and nothing beyond.
- Technical rights cut to that scope, not yours.
- Human approval on actions that commit you: sending, paying, deleting.
- A complete log: what it read, decided and did, readable without us.
- A ceiling on spend and on calls, because a model's cost has to be watched.
- A switch: cutting the agent off must not require anyone's permission.
An agent does not crash, it gets it wrong
The log and the switch are not a lawyer’s precautions. An agent goes wrong differently from a program: it does not crash, it does something plausible and false. Without a trace, the error surfaces weeks later, and without a switch it carries on while you are looking for where it came from.
What we will not build
We turn down certain uses, and it is worth saying so here. Anything that decides on its own about an entitlement, an individual price, a hiring or a refusal falls under the European regulation on artificial intelligence, and calls for a very different framework from the one this page describes.
If your need stops at getting the right answer, an assistant costs less and is easier to supervise. And if the steps are always the same, a connection between tools will do the job without any model having to decide anything.