AI agents
Some tasks resist conventional automation
Hand over to a machine the tasks that require reading, understanding and deciding.
They come back every day, they are repetitive, and yet no script can do them: you have to read a message and understand what it asks, connect two pieces of information that don’t have the same name, or choose the next step among several possible ones.
The result: someone does it by hand. That is expensive time spent on dull work, and it is exactly where mistakes pile up as soon as the workload grows.
An agent is a program you entrust with a decision
An automation follows a path written in advance. An agent, on the other hand, is given a goal, tools and limits, then chooses for itself how to go about it. That is what allows it to handle cases nobody had foreseen.
I design the agent, connect it to your tools, and leave the final say to a person wherever a mistake would be costly: nothing irreversible happens without approval.
- A scope written in black and white: what the agent can do, and what it will never do on its own
- Human approval before any irreversible action: sending, publishing, paying, deleting
- Connected to the tools you already use, without forcing you to switch
- Every decision is logged: you can always find out why the agent acted the way it did
- Your data stays yours: together, from the start, we choose which model to use and where it is hosted
What I build them with
The choice of tool depends on what the agent has to do, not on the latest trend.
Hermes Agent
My go-to tool when you need an autonomous agent, up and running quickly, without starting from scratch.
N8N
When the agent mainly needs to coordinate existing tools, with steps that are visible and easy to change.
Python
When your business has rules too specific for a standard tool.
Language models
To read, sort, write or extract information. Hosted by a provider, or on your own servers.
MCP
The protocol that gives an agent controlled access to your data and tools, within precise limits.
Your tools
Email, CRM, invoicing, databases: the agent works where you already work.
How it works
1. We start from a specific task that someone does by hand today
2. I measure how much time it costs, and how much of it an agent can really take on
3. I build the agent and first run it in dry-run mode: it suggests, but does not act yet
4. We hand it actions one at a time, starting with those that can be undone
Still have questions?
I answer all your questions.
-
How is it different from an automation?
An automation follows a path you have described: if this, then that. It works perfectly as long as the cases are known, and it gets stuck as soon as an unexpected case comes up.
An agent is given a goal and chooses the path itself. It is more flexible, but that is also why it needs written limits: without them, it creates more risk than it saves time.
Many tasks don’t need an agent. When an automation is enough, I tell you so: it costs less, both to build and to run.
-
Does my data go to an AI provider?
Only if you decide so. This is settled during scoping: depending on how sensitive the data is, we choose a provider that commits in writing not to reuse it, or a model installed on your own servers.
In every case, what may leave your premises is set down in writing before development begins.
-
What if the agent makes a mistake?
We assume from the outset that it will. That is why irreversible actions go through human approval, and every decision is logged: a mistake must be spotted and fixed right away, not three weeks later.
-
How much does it cost?
Building it is quoted after scoping: that is what shows whether the task justifies an agent.
On top of that comes a running cost, which depends on the model used and the volume processed. It is estimated during scoping, so that the first bill is not a surprise.