On September 29th, I spoke in the closing plenary of RONOG 11, the Romanian Network Operators Group meeting hosted by InterLAN at the Radisson Blu in Bucharest, with a talk called "Tools and a Recipe Is All You Need".
I started with a question for the room: any other dinosaurs who used to program in assembly? Programming started with ones and zeros, moved to Fortran, then to Python and Java. In 2026 it has reached our own language. And I don't mean chatbots. I mean agents like Claude Code or Codex, which you program in plain English or Romanian, and which then follow your instructions on your own machine.
Stop being the copy-paste department
Most of us still use AI like a consultant on the phone. You describe the problem, it gives you an Excel formula, you paste it, you paste the error back, it apologizes and gives you a new formula. The model is the brain, and you are its hands. An agent gets its own hands: it can read and write your files, write and run code, and try, fail and correct itself until the task is done. It is not smarter than you, but it reads a hundred-page document in seconds and can run many sub-agents in parallel.
"If you can explain a task to another human, and that human can do it without you, then that task is a great candidate for an AI agent."
Tools and a recipe
This is where the title comes from. When you teach an agent a task, it writes a skill: a recipe that mixes instructions in human language with real code. You run it today, you fix what went wrong, and tomorrow it runs better. The tools come from connections. An agent with no access to data can't do anything, and MCP is the Lego piece that plugs it into other systems. Earlier that day, PeeringDB had announced their MCP server, which was a perfect example: instead of waiting months for a new button in their interface, you ask your own agent for any report you want, straight from their data.
"What if the AI makes a mistake?"
This question came from the audience, and it led to the liveliest part of the session. My answer: you already know how to handle this, because your human employees make mistakes too, and the consequences land on your shoulders either way. The job is to catch mistakes before they reach the real world. With AI you do it with checks that always give the same answer: a balance sheet must close on zero, a bank account number has a checksum, a ledger of inputs and outputs shows the one record that went missing. Don't accept "yes boss, I'm sure". Ask for the report, the reasoning and the proof, and let other agents check it. When I write software today, I ask the agent to think of every way the app could fail and write a test for each. I end up with a thousand tests, where I used to write twenty.
Watch the session
I closed with a simple ask: an AI agent is like a new colleague who is great with computers but knows nothing about your business, so teach it. Install Claude Cowork or OpenAI Codex, it's about 20 dollars a month, and give it a real task. Thanks to InterLAN and the RONOG program committee for having me, and to everyone who kept the discussion going at the party that evening.