Let me make a case I do not hear made often enough.
For two years, enterprise AI has had one volume setting. Agents. Systems that go off and do the work end to end while you watch the result. The ambition is real, and a large share of it will pay off. I am not here to argue against it.
I am here to argue for the half we keep leaving out.
Something quieter has been happening after the pitch. Autonomous deployments are being scaled back. Pilots that looked sharp in a demo did not survive contact with a function that has to answer for the result. The easy reading is that the agents underdelivered. I think that reading is incomplete.
The agents did not fail at capability. They ran into accountability.
Capability was never the wall
In a lot of work, full automation is the right answer. High volume, reversible, low consequence: an agent should own it. Nobody wants a person hand-sorting ten thousand records or triaging routine tickets one at a time. That work is going to agents, and it should.
But some decisions carry a name and a number. The person who signs the contract, approves the budget, or commits to the forecast owns that outcome to someone above them. When it goes wrong, no one in that room accepts "the agent decided." You can delegate the analysis. You cannot delegate the accountability. That is not a weakness of the technology. It is how consequential work has always been governed.
So the useful question is not "agent or not." It is which decision class you are in. Reversible, routine work runs well unattended. Consequential, judgment-heavy decisions with a named owner want a different posture. Both deploy. Often in the same company, sometimes in the same workflow. This was never a contest.
The posture no one named
Here is the gap. Everyone has language for the agent. Almost no one has a confident name for the other half. When supervision does come up, it gets filed as a consolation prize. Human in the loop. The thing you settle for until the agent is ready.
That framing is wrong, and I think it is about to get corrected.
Supervision is not a weaker form of automation. It is a deliberate design choice that matches how accountable work actually gets done. The expert does the legwork and advises you through it. You keep the decision. Done well, it is not friction with good manners. It is the difference between a tool that hands you a finished answer you have to second-guess, and one that shows you the work so you can stand behind it.
The cleanest model for it is not in software. It is on a golf course. It is a caddie.
A good caddie has walked the course a hundred times and reads it better than the player ever will. They carry the bag, study the lie, judge the wind, and tell you the line. They hand you the club. Then they step back, because the swing is yours and so is the score. Expert, and never presumptuous. A caddie makes the player better without once taking the shot.
That is the posture. Call it the Caddie. An AI that does the analytical legwork and advises you through the work, then hands you the decision and lets you make it. Not a system that plays the round for you. One that makes sure you play it well.
What a caddie does
A caddie built in software does three things, and all of them live inside the work rather than in a separate stream of advice.
It proposes. You never start from a blank page. The first draft is the advice, with the reasoning visible in the draft itself. The line is read by the time you step up.
It shows consequences before you commit. Make a change and you see what it does before it is final, not after. You read the lie before you swing.
It observes. It points at the thing you did not ask about, the detail that will cause trouble later. It flags it once, plainly, and lets it go if you proceed. It does not nag, and it does not tell you what you must do.
Then it defers. You hold authority over the decision and over when the work is done. The caddie hands you the club. You take the shot.
None of this is exotic. It is most of what a good analyst, associate, or advisor already does. The shift is that AI can now hold that role at speed and scale, across far more of the work than a human bench ever could.
Use both. Know which is which.
Agents and caddies are not rivals. They answer to different kinds of decisions. An agent is the right call when the work is high volume and a wrong move is cheap and recoverable. A caddie is the right call when a person has to own the result and wants to understand it before they sign. A serious system runs both, each where it earns its place. Anyone selling you one answer for everything is selling you the demo, not the deployment.
The last two years treated autonomy as the destination and supervision as the waiting room. That was the mistake. They are two tools. The whole skill is knowing which decision is in front of you.
Where this goes
I do not think supervised collaboration stays unnamed for long. The pattern is already showing up in how careful buyers talk. They are not asking for a system that acts for them on a decision they own. They are asking for one they can interrogate, push on, and approve with their eyes open.
The last two years taught us what to automate. The next few will teach us what not to. That is the harder lesson, and the more valuable one.
So here is the case, in one line. Automation was never the goal. Good decisions are. For the ones that carry a name and a number, you do not want a system that plays the round for you. You want a caddie. It reads the line and hands you the club. The shot is yours, and it should be.
One admission before I go. I know how this plays out. "Agentic" went from a sharp idea to a marketing category in about a year, and "Caddie" will make the same trip. So before anyone else names it, I am planting the flag before the rest of the field shows up. You will see Caddie AI the way you saw Agentic AI. I would rather call it than chase it.