The fact: on 17 September 2026, Anthropic published its R&D automation index. Claude now drives 26% of the work that builds the next Claude. Around 30,000 agents were running in parallel on its internal platform in August. And out of more than a billion monitored agent decisions, only one in 47,000 was blocked.

The Codito take: the figure that matters is not the 26%. It is that Anthropic publishes a block rate. Measuring what your agents produce is easy; measuring what you prevent them from doing assumes that someone in-house has drawn the line. That is the whole distance between a system and a pile of tools. The day before, on 16 September, Collibra and the Harris Poll surveyed 306 data and AI decision-makers: 51% spend a significant number of hours checking and correcting by hand what their autonomous agents produce, and 76% have hit a critical obstacle between pilot and production in the past twelve months. This is not a weakness of the models. It is the absence, in-house, of people able to put the guardrails in place.

What it means for you: Anthropic does not run 30,000 agents because it has a better model, but because its teams know how to build them and keep them in check. This week, take a single process you have already automated and write down in one sentence what triggers a human check. If nobody around the table can answer, your issue is not the tool, it is the skills.

88% of AI pilot projects never make it into production (IDC / Lenovo). This failure is often blamed on the technology. The truth is simpler: a pilot is carried by one person, often the CEO or an enthusiast in the team, whereas production is carried by a group. What is missing between the two is rarely a tool. It is a skill that nobody has passed on.

That is precisely the work of the Premium AI Partnership. Every month, two hours with you to stay on course, and two hours with your teams to pass on the skills. The training is not a generic course on AI in business: it is tailored to your tools, your processes, your sector. And during the session, we do not explain how we would set up an agent, we set up one or two with your staff. They leave with the agents running and the skills to add more on their own.

I owe you the boundary, because it is part of the deal: complex agents connected to your CRM, your ERP or your document management system are not set up in a workshop. Those fall under Custom Development, and we tell you so beforehand, not afterwards. The goal has never been to make you dependent on Codito. It is to make you self-sufficient with AI.

Codito is certified by Anthropic. Your teams therefore train on the same building blocks as those running the 30,000 agents I mentioned above, at the scale of your business.

Or book 30 minutes to talk it over in person

I told you about GEO right here two weeks ago. Since then, I have dug deeper, tested it live, and turned it into a thirteen-minute masterclass published on Friday.

The starting point: since 22 July 2026, one Google search in two in France gets an answer written by an AI, before the first link. If you are not in that answer, you do not exist in half of the searches that concern you.

What surprised me most while preparing it was visitor behaviour: for the same traffic, someone who reaches you from an AI contacts you four times more often than someone coming from a Google link. The AI has already done the sorting for you. Second surprise: backlinks, the backbone of traditional SEO, have almost no effect here. I show where it really plays out, live, on a law firm in Toulouse.

The interactive guide is free, with no email and no form: the diagnostic to paste into ChatGPT, the simulator, the seven-step protocol and the source of every figure.

And if you do only one thing after reading this: open ChatGPT and type the question a client asks just before finding you. Not your name, their problem. Then look at who gets cited. In two minutes you will know where you stand.