For most of my career there were two complaints you could make, and both of them were usually true.
From the engineering side: my product partner isn’t strong enough. They can’t hold a clear direction, and they don’t carry the business outcome the way the role is supposed to.
From the product side: engineering only takes care of executing their own piece. Nobody over there owns whether the whole thing works for a customer.
I made the first complaint for years. I meant it every time.
What I didn’t see then is that both complaints were also permission slips. If direction wasn’t formally my job, then the absence of direction wasn’t my failure. And if the outcome wasn’t formally engineering’s, a product lead got to be the only person in the room carrying it. Each grievance came with a small exemption attached, and the exemption was the part we didn’t examine.
Both are being taken away right now.
The era of the general manager
What made those complaints possible was the division of labor itself, and that division is dissolving.
You can see it in hiring. A company opens a leadership role, and the people seriously considered include someone who spent a decade as an engineering lead and someone who spent a decade as a product lead. The listing didn’t say which background it wanted, because it no longer needs to. The opportunity stopped being defined by function.
You can see it in reporting lines too. Engineers reporting to a product lead used to be unusual enough to comment on. Now it’s one ordinary arrangement among several, and the classic pair of an engineering manager beside a product manager is no longer the default shape of a team.
The old structure needed that pair. Two people, two vantage points, and enough coordination between them to keep a growing team pointed the same way. In practice the line was never as clean as the org chart claimed — any engineering lead reading this has made hundreds of product calls that were nominally somebody else’s. But the pair was structurally necessary, and at a certain size it no longer is.
What arrives in its place is the general manager. Not a title to go get. A shape: someone who owns the whole thing, and who is now measured that way whatever their business card says.
What’s being measured
When I set the titles aside and look at what these roles are measured on, three things keep surfacing.
Product sense. Technical sense. Collaboration.
Underneath all three is the same thing, which took me a while to see. They are all forms of taste.
Not taste as decoration. Taste as accumulated judgment about what’s good — the kind nobody can hand you and no rubric contains. Product sense is taste about what to build. Technical sense is taste about how a system should be shaped. And collaboration is how you get the taste you didn’t earn yourself, by learning from what other people failed at.
Because that’s where taste comes from. Taste is built out of failures you lived through. The things you shipped that didn’t land. The architecture that looked elegant and became unmaintainable in eight months. The hire that didn’t work. Your taste has the exact shape of your own history of being wrong, and nobody else’s history produces it.
That one fact explains most of what follows.
Product sense
Every other quality here has a proxy. Technical skill has credentials and code. Product sense has none. It shows up in the product or it doesn’t exist anywhere.
Which is why it’s the hardest of the three to fake and the most valuable to have right now. When building was expensive, choosing what to build was one decision among many. Now that building is cheap, choosing is close to the whole job.
The quality underneath it is a willingness to be wrong in public. Not wrong in a retro. Wrong where a real person does something you didn’t predict, and you have to sit with why you didn’t predict it. That’s the correction signal, and there is no substitute for it — which is also why you can’t build product sense by thinking. Frameworks tell you where to look. They don’t tell you what to conclude.
This is why doing your current job harder doesn’t develop it. In most engineering roles someone else owns the product call. You can ship for years and never learn whether your judgment was good, because your judgment was never the thing under test.
So the way to strengthen it is to own the calls somewhere. Starting a company is the expensive version. A side project is the cheap one, and it’s the same loop at a different price: build something small, put it in front of real users, and let their behavior tell you where your instincts were off.
I say this as someone who was comfortable making product decisions long before I started building something of my own. It still took time. I had to learn, get things wrong, and build the taste for it. Being willing to make the call and having good judgment about the call turn out to be different skills, and the distance between them closed slowly.
Technical sense
The worry I hear most is that technical depth stops mattering once agents write the code. What I’m seeing is closer to the opposite.
The hard problem in running agents turns out to be information architecture, not agent count. How do you scope context by domain so an agent loads what it needs and nothing else? If agents are stateless, where does persistent knowledge live, and who keeps it current? What credentials does each one get, and what should it never be able to reach?
Those are the oldest arguments in engineering. Decomposition. Where state lives. Access control. The context window didn’t invent them, it made them urgent. Security is the sharpest case — the tooling has no granular access control yet, so teams isolate agents outright, onboarding and offboarding them the way they would a hire.
Here’s the part that requires depth. You can’t judge the quality of output in a domain you don’t know. An agent that stalls after ten minutes is an annoyance. An agent that runs for hours on a bad structure does damage; I know of one that filled a repository with megabytes of build artifacts before anyone noticed. Someone without technical depth sees that something went wrong. They can’t see why, and they can’t tell a real fix from a plausible-sounding one. From outside, the fix and the excuse look identical.
That diagnosis is technical sense, and it’s the harness you’re building whether you call it that or not. You strengthen it by keeping your hands close enough to the work to still diagnose it. Review specs and acceptance criteria rather than diffs, because the spec is where the judgment now lives. And stay current on the fundamentals, because they are what let you evaluate an answer you didn’t produce.
I’ve been sharing my learnings of agent harness through AI-Native Org series.
Collaboration
Influence matters less than I was taught. Not all influence, but the kind most of us trained on — persuading other people to build the thing you believe should exist. That mattered when building was expensive. Now you can build much of it yourself.
Alignment on direction is a different matter. Building got cheap. Agreeing on what to build did not.
And if taste is built out of the failures you lived through, then your taste has the shape of your failures and the holes of everything you never had to survive. You can’t see around your own experience by trying harder at it. AI doesn’t solve this either. Harnessed well it connects dots impressively inside the material you hand it, but at the level of big direction it’s weak. It has the material and none of the living.
So the correction has to come from someone whose failures weren’t yours. My teams have always worked the same way: implement something, put it in front of each other, take the feedback, iterate. Not because it ships faster, though it does. Because feedback is the form taste travels in.
Ask people what they think in the abstract and you get opinions. Put something concrete in front of them and you get their taste — the specific thing they notice, the thing that bothers them that you walked past. Usually because they’ve been burned by it before and you haven’t. That’s the same loop as product sense, pointed at colleagues instead of users, and it’s how a group reaches a decision better than anyone’s individual judgment in the room.
The quality underneath is humility, and a real appreciation for difference. Everybody has their own genius, shaped by a history you didn’t live. If there were one right answer about taste you could look it up. Because there isn’t, more than one set of eyes is the only correction available.
No more exemptions
This shift was already underway. AI is accelerating it by making the boundaries between functions cheaper to cross and harder to defend.
What matters more now is the judgment to see the whole, the depth to know where your own judgment is weak, and the humility to borrow from someone else’s.
It’s whether, together, we are willing to build the tastes to own the whole thing — and whether I’m willing to own my part of whatever is missing.