Where this goes in the next twelve months
Someone asked a fair question under the last piece. With all that experience, how do I see this playing out over the next year? Here is my honest guess, no hype and no doom, from someone who has to actually run the stuff rather than sell it.
The capability stops being the story
Not where the action is.
The tools are going to get a lot more capable, and quickly. That is the safe prediction and also the least interesting one. Once an agent can do most of the hands-on work competently, the question stops being "can it do the task" and becomes "what is it allowed to touch, and who is on the hook when it is wrong". That is not a capability question. It is an old, boring management question wearing a new coat.
The bottleneck moves to governance
Software catching up on discipline the trades sorted decades ago.
So the interesting action over the next year is not in the models, it is in the plumbing around them. The businesses that get ahead will not be the ones that give an agent the most freedom. They will be the ones that work out how to let it do far more while keeping a hard human line on anything you cannot take back. More speed, same accountability.
That is a solved problem in industries that have always had to live alongside things that can hurt you: aviation, finance, construction. It is mostly unsolved in software, which has never really had to. The next year is software quietly catching up on discipline the old trades sorted out decades ago.
A lot of people will get burned first
And it will be depressingly predictable.
Before that settles, plenty will get burned, and the stories will all rhyme. Someone gives an agent standing access to a live system because wiring up a proper approval step was more effort than they had that afternoon. It runs fine for weeks. Then it does one confident, wrong thing at machine speed, and by the time a human looks up it has done it a thousand times.
None of those stories will be about the AI being evil. They will be about access nobody thought hard about. The lesson is an old one: the problem is never the tool, it is what you let it reach.
The skill that quietly becomes rare
You can only supervise a job you could do yourself.
There is a quieter shift underneath all this, and someone in the comments on the last piece put it better than I could. You can only supervise a job you could do yourself. The instinct for when the tool is talking rubbish does not live in the tool, it lives in the person, and it is built from years of doing the work the slow way.
Over the next year that instinct goes from common to genuinely valuable, because the people who have it can let an agent fly and still catch it, and the people who never built it are just trusting a confident stranger. If I were advising anyone starting out, I would still tell them to learn the craft properly. Not so they can do it by hand forever, but so they can tell when the machine is lying to them.
The question that matters
Who signed off on this change.
So my boring prediction for the next twelve months is that the most important sentence in a lot of businesses becomes "who signed off on this change". Not the cleverest model, not the flashiest demo. Just a clear, human answer to who is accountable and how you undo it.
The companies that can answer that will hand more and more to their agents with a straight face. The ones that cannot are in for a very exciting year, and not the good kind of exciting.
This is how we build
More nerve with your systems, and the same accountability underneath. If that is how you would want yours run, get in touch.
Drop me a line