Years ago, before children and before the part of life where you start counting risks, I rode Italian motorcycles. Ducati, MV Agusta, anything with a twin engine, too much torque and a sound that made people at traffic lights turn their heads. It is a hobby that kills a certain number of its enthusiasts every year, and I decided early that if I was going to keep it I would take every advanced and sports riding course I could find. Most of what they taught has faded into reflex. One thing did not.

The instructor put it plainly on the first morning. The human brain was built to take in the world at the speed we walk and run. Some people manage a little more, some a little less, and almost nobody manages what arrives at 120 miles an hour. At that speed the road in front of your wheel is already the past. By the time you see the gravel in the corner you are on it. So the whole discipline of riding fast is a discipline of not looking there. You look as far ahead as the road allows, you read the corner before you enter it, you decide where you will brake before you need to, and you keep, at all times, a plan for the thing that has not happened yet. A rider who gets on a sport bike unprepared and stares at the tarmac in front of the fender is not riding. He is waiting to find out what the road has decided for him.

I have spent the last two years studying AI and building agents with whatever coding tools were current that month, and I have written several books about what that does to the product manager and the team around him. Recently I built the same small product twice, once on my laptop with a coding agent as the builder and a frontier model as its supervisor, once in the cloud with a different pair, to learn what happens to the work when implementation takes hours instead of weeks. What happened to me was the first morning of that riding course all over again.

The agent produced a pull request roughly every forty minutes. Not every forty minutes of my attention, every forty minutes of the clock, through lunch, through meetings, through the evening. In six days the record of decisions I had to make grew to forty-four numbered entries, each an answer to a question the agent had parked because nothing in the instructions covered it. I did not have forty-four decisions in me at that pace. Nobody does. I was riding at 120 with my eyes on the road in front of the wheel, answering the question that had just arrived, and the product was worst exactly where I had been looking hardest.

The dashboard told me otherwise. Five days in a row the nightly run finished green: fetched, gated, scored, ranked, reported. On the sixth day a blind sample of postings that I had judged by hand showed that thirteen of fifty-eight good ones had never been read at all. They sat in a queue for descriptions that had not been fetched, silently, while every light on the panel stayed on. I had seen precisely as far as the headlight lit, and at that speed the headlight is not far enough.

The instinct, when the machine is that fast, is to try to be faster. Answer quicker. Watch closer. Read every diff. That instinct is the pothole. No human matches the velocity, and the effort to match it puts your eyes exactly where they should not be: on the line of code that just appeared, the question that just arrived, the test that just failed. The technique is the one from the course. Everything that matters moves to before the ride, and while the ride is on you look at the exit.

Before the ride begins there is a question that the riding school asks and that most teams adopting agents skip. Speed is chosen. Nobody rides at 120 through a town, not because they lack skill but because no amount of skill makes it survivable there. The track exists so that speed has run-off, a known surface, marshals and flags. In "Agentic AI for Busy Product Managers" this is the suitability test and the go/no-go decision, and it comes first for the same reason the speed decision comes first: the question is not whether I can keep up but whether this is a road on which this speed can be survived. A task with irreversible actions, thin context and no way to check the result is a town at 120. The right answer is not a better rider. It is a different road, or a slower ride, or no ride.

Then the map. On the sighting lap you learn the corners, and by the time you go fast you are not discovering the track, you are executing a plan you already hold. The books split the plan into two documents for a reason that only makes sense at speed. The Human Brief is the horizon: what the ride is for, what done looks like, what is not on the route at all. The Executable Brief is where your eyes are while you ride: the constraints the agent can check itself against, the behaviors that are never allowed, the acceptance tests that say the corner was taken. The distinction matters because of a second lesson from the course, one that took a few near misses before I believed it. The bike goes where you look. Stare at the pothole and you hit the pothole; look at the guardrail you are afraid of and you will meet it. Vision on a motorcycle is not observation, it is steering, and the instructors drill it because it is the opposite of instinct. Instinct looks at the danger. Training looks at the exit. The agent goes where the spec looks. A brief that describes the obstacles in loving detail, every failure mode, every thing to avoid, steers the agent into them with the same fidelity a frightened rider steers into the guardrail. The Executable Brief has to be written as the line through the corner, not as a catalogue of the things beside it.

Then the markers. On a fast lap you do not decide where to brake in the corner. You decided it on the sighting lap, you chose a marker, and you brake at the marker whether or not it feels early, because in the corner your judgment is the last thing you should trust. The approval moment is that marker. It sits at the irreversible action, sending, submitting, deleting, merging, and it is set before the work starts, not negotiated while the agent is mid-corner with a green test suite and a plausible reason. The cousin of the marker is a habit every instructor insists on: cover the brakes, two fingers resting on the lever through the whole fast section, so that the stop is always available without being applied. That is the autonomy boundary, and it is not a paragraph in a document. It is an allowlist in code, a list of the actions the agent cannot reach whatever it is told, and in the laptop project it is the only rule that was never broken, because it was the only rule that did not depend on anyone being awake.

And fuel. The one that sounds trivial until you are on the shoulder. The agent does not tire and does not stop for the night; the budget does, the tokens do, and the human attention certainly does. A ride with no budget for how far it may go, no alert when the burn rate climbs, no ceiling on what one run may consume, ends the way every ride with no fuel plan ends, somewhere inconvenient, at the worst possible moment, with the destination in sight.

Here is where the analogy breaks, and the honest version of it is more useful than the seamless one. On the bike the rider is the one moving. Adrenaline, peripheral vision, the feel of the front tire through the bars, all of it is on the rider's side, and a good rider can improvise a little when the plan runs out. With an agent the human is stationary and the machine moves, at a pace you cannot feel in your hands, and your instincts are not merely useless but inverted. The thing you want to look at, the diff that just landed, the question that just arrived, is the pothole. There is no feel for the front tire. There is a log.

Which is why the whole discipline moves to before the ride. On a motorcycle perhaps a third of what keeps you alive is decided in advance and the rest is technique in the moment. With an agent the proportion flips. The suitability decision, the two briefs, the approval markers, the boundary in code, the fuel plan, and the evaluation set that tells you afterwards whether the corner was actually taken or only felt like it, are not preparation for the work. They are the work. What remains in the moment is the one thing a stationary human can still do at any speed, which is to look at the exit and refuse to look at the pothole: to answer the parked question with the destination in mind rather than the diff, and to trust the markers you set when you were calm.

The evaluation deserves its own sentence, because it is the part of the lap that has no equivalent in the rider's body. On the bike you know whether the corner went well. With an agent you do not, and the dashboard will tell you it did. The eval is the only instrument that reports what the road decided rather than what the panel displayed, and it is the reason the thirteen unread postings became a number on a board rather than a feeling that the list had gone quiet.

I wrote "Agentic AI for Busy Product Managers" and "The Agentic AI Team" while riding this particular road, at a lower speed than the laptop project reached but on the same surface, with the same agents parking the same kinds of questions. They describe these artifacts, the suitability test, the two briefs, the approval moment, the autonomy boundary, the eval that settles arguments, as the product manager's deliverables in an age when implementation is no longer the slow part, and they describe them from the saddle. What the higher speed added is a sharper feeling in the hands, or rather a sharper sense of their absence. The books were right about what to write down. Riding faster showed me how little of you is left in the moment once the machine is moving, and how completely the ride is decided by what you did while it was still parked in the garage.

The bike goes where you look. So does the agent. Look at the exit, set your markers before the corner, keep two fingers on the lever, know how much fuel you have, and above all decide, before the engine starts, whether this is a road you should be riding at all.

Part of the Product Management series: how AI is changing what product managers actually do. The frameworks named here, the suitability test and go/no-go, the Human and Executable Briefs, the approval moment, the autonomy boundary and the eval, are developed in "Agentic AI for Busy Product Managers" and "The Agentic AI Team".