Steer, Don't Row

Once AI takes over execution, your job shifts from rowing to steering. What steering is, how the abstraction rises, and who owns the outcome.

Work will look more and more like steering than rowing. The metaphor comes from Tara Seshan, the OpenAI product lead behind Codex and ChatGPT Work: agents do a lot of the rowing, and your role increasingly becomes pointing the ship in the right direction.seshan

After reading, you should be able to answer:

• Where the line actually is between rowing and steering
• Why the abstraction level of steering keeps climbing
• Why accountability and judgment have to stay on the human side

This shift isn't caused by any single product; it's caused by what AI can do. In the past, writing a line of code and pressing Tab was rowing. Now you say a sentence and the code is already written — your job becomes deciding whether that code should exist, for whom, and where to go next.

What rowing is, what steering is

Rowing is execution: writing code, building spreadsheets, organizing materials, tracking progress, turning ideas into deliverable artifacts. Steering is direction: deciding where to go, why, what to do first, when to stop, and what you will absolutely not build.

The line between them isn't skill; it's the question being answered. Rowing answers "how" — it has a standard answer, it can be verified, and so it can be handed to a tool. Steering answers "where, why, and whether it's worth it" — there's no standard answer, it requires judgment, and for now it has to stay with a person.

AI is turning more and more of "how" into "one sentence." So the time you have for rowing shrinks, and the places that need steering multiply. It doesn't happen all at once; each piece gets handed over one at a time. First code completion, then whole functions, then prototypes, tests, docs, and then entire features.

The abstraction keeps rising

Once the rowing goes to agents, steering itself gets more abstract. The lowest level is "I wrote this line and pressed Tab" — that's rowing. One level up is "have the agent do this whole task, I'll accept it" — that's starting to steer. Higher still is directing bigger things: setting goals, choosing strategy, deciding whether to build at all. The top level is the goal itself: what this product should become in three years.

Rising abstraction doesn't make your life easier; it moves you further from execution and closer to judgment. At each level you have to relearn one thing: how to express a direction clearly enough for an agent to execute, and how to spot problems in what the agent produced instead of fixing it yourself. That's harder than doing it by hand — because you can no longer dodge the judgment by saying "I'll just do it myself."

Steering requires opinionated calls

Part of steering comes from data, but a large part is making opinionated calls — even a kind of positive commitment to a future: I'm choosing to make the product look like this, not because every other strategy is impossible, but because I want the world to move in the direction I'm pushing.

That distinction matters. Data tells you how things are; it doesn't tell you where the world should go. When everyone has the same tools and can produce the same things, what separates people is point of view, taste, and judgment — not who rows faster. Products are the same: they're more like films than real estate. Throwing money at them doesn't guarantee value; they need authorship and a point of view.

Someone has to own the outcome

When role boundaries blur — engineers doing PM work, PMs shipping prototypes — the real danger isn't messy division of labor; it's that nobody owns the outcome. Someone has to be directly responsible for whether the product is used, wanted, high quality, and effective — whether that person is an engineer, a designer, or a product manager. That role is the DRI (Directly Responsible Individual).

The actual work can be claimed by anyone with interest and ability, but the team shouldn't argue about role boundaries; it should focus on getting results. When a call goes wrong, asking who owns it is the cleanest way to tell roles apart — and this line matters more, not less, in the AI era. Once things move faster, anything nobody owns degrades just as fast.

Steering as a group

Today, working with agents is mostly one-on-one: you and your agent, at most with a few sub-agents spawned underneath. But that stays cut off from the agent work of your colleagues — teams used to paste screenshots of agent threads into chat to prove how a number was reached. That isn't the most natural way to collaborate.

The future looks more like a multiplayer game: a group of people steering a group of agents together. Your agent does an analysis, and my agent checks whether it's trustworthy; everyone reviews the work in progress and gives feedback. Bringing other team members into your collaboration with agents is the direction steering moves as it becomes a team activity.

How to check whether you're rowing

Take whatever you're doing right now and ask three questions. Does this have a standard answer? If it does, can it be handed to an agent? If both answers are yes, you're rowing — even if it used to be your core skill. The work that's actually worth your time is the other kind: no standard answer, requires judgment, and the blame lands on you when it goes wrong. If you find most of your day still going to "how," the tool isn't the problem; you just haven't handed the ship to the helmsman yet.

A practical rule of thumb: if an agent can do it, you shouldn't keep doing it; you should do what the agent can't — set direction, make trade-offs, own the outcome. That sounds light but isn't: rowing gives you a sense of completion, steering doesn't. The payoff of steering only shows up much later, but it's the only place your lasting value lives.