- Not a chatbot any more When did your AI assistant stop being a chatbot? The same model is a low-risk adviser and a high-risk executor, and the difference is one connected tool. An inventory you can finish this week: what the agent may read, change, send and delete, and what each costs when it is wrong.
- Breaking the cascade How does one small mistake reach five systems, and where do you break it? In a chain of automated steps, an uncertain output arrives at the next step as a trusted input. Chain depth, retry budgets, idempotency, a re-check of the critical fact before anything irreversible, and operations that can be compensated.
- Where the model runs Does running the model yourself actually make it safer? Bringing a model in-house changes who controls the infrastructure and where data goes. It changes nothing about prompt injection, over-broad permissions or unsafe tools. Four questions that decide the choice, and the risks that survive it.
- The blast radius How do you keep one wrong decision by an agent from becoming an incident? An agent is safe because of what your infrastructure physically lets it do, not because the model is good. What to bound: scope per object, a ceiling on volume, an expiry on every permission, confirmation before the irreversible, and a way back.
- Code is not the advantage Code stopped being the advantage. It did not stop being the constraint. Software is written faster than it has ever been, so the scarce thing moved. What an owner runs short of now is not engineering hours but a dependable answer to how their own system works. What that shortage costs, and how to end it.
How these are written
- One transition per piece. Each article takes a reader from "this is probably a problem for somebody" to "this is a problem for me and here is the first thing I will do about it". If it cannot name that move, it does not get written.
- The cost of waiting is made explicit, and only with things that can be checked: what accumulates, what it looks like in three months, what the next big change will make of it.
- Every article ends in your hands, not in a warning. A sequence, a checklist or a way to test your own position, and a first step small enough to take today.
- Every invented example says so. There are no customer stories here, because there are no customers to write about yet.
- A date on every piece. An argument about a moving subject is read differently when you can see how old it is.