Recently, I've been working intensively with coding agents, and I made a surprising discovery.
AI excels at getting things to 80% or even 90%. Give it a task, and it quickly produces something that looks good. But "looking good" isn't the same as being valuable. Software engineering has no objective standard of perfection—zero-bug code that nobody uses is worth zero. Direction matters far more than execution.
Our team shipped more code in the past month than in the previous six months. That sounds impressive, but it actually makes me anxious—how much of this code is truly solving user problems? How much of it is just "because we could"?
Two States
Lately, I've been oscillating between two states.
When the hypothesis is clear, I can't stop. I know what needs to be validated, and the coding agent helps me rapidly build the relevant components and run end-to-end experiments, compressing the entire pipeline down to hours. In a single day, I can iterate through two or three hypothesis corrections—ideas are confirmed or refuted within hours. The rhythm is addictive.
For example, last week we were testing a hypothesis: would customers pay for AI-generated ops reports? In the past, validating this would require pulling in a product manager, a frontend dev, and a backend dev—two to three weeks just for a demo. Now, working alone, I spent most of a day building a working prototype and sent it to three customers for trial that same afternoon. The results: two said "nice, but not worth paying for," and one said "this would be great if it could integrate with our monitoring system." Half a day to validate a hypothesis, saving perhaps three weeks of detours.
But there are also days when I sit at my desk with no idea what to do. Without anything specific to validate, I dig up previous projects and have the agent polish this or fix that. It looks like work, but I know deep down I'm just idling.
Both states use the exact same tools. The difference is entirely in whether there's a clear question in my head that needs answering.
The Funnel Is Broken
Traditional product management has a classic concept called the ideation funnel.
There's an initial "fuzzy front end" stage where you generate a lot of ideas, then filter them. Why filter? Because implementation downstream is too expensive. A product going from concept to launch typically takes three to six months of development. Since implementation is costly, the front end has to be strictly gated.
That logic no longer holds.
An MVP can be built in a day for a few hundred dollars in API costs. An idea can go from birth to being in users' hands in a week, or even days. Implementation is no longer expensive, so the premise for filtering disappears. Ideas used to be cheap and execution expensive; now it's the opposite.
You can imagine what this means: previously, a company might only validate ten product ideas a year because each required massive manpower and time. Now the same team can validate a hundred. But here's the question—do you actually have a hundred ideas worth validating? Most teams can't even come up with ten high-quality hypotheses.
Suppressed for Too Long
The work environment of the past has actually been suppressing the impulse to "come up with ideas and validate them."
Most people were expected to execute. If you had too many ideas, you'd likely hear: "What's the point of thinking so much? Just do your job." In an era of limited resources, everyone fought for execution resources, not ideas. So many people's ability to generate hypotheses has been atrophying.
Now the agent era has arrived, execution is suddenly unrestricted, and we have a productivity surplus. But this surplus productivity doesn't know where to go.
I've observed an interesting phenomenon in our team. Those who were previously "suppressed"—people with lots of ideas but no resources to execute—are now thriving. They've accumulated a wealth of unvalidated hypotheses in their minds, and agents have given them an outlet. Meanwhile, those who focused solely on execution suddenly find their core competency being replaced by agents, without enough "idea reserves" to fill the gap.
Impact on Different Roles
Engineers are the most directly impacted. The barrier to writing code has dropped dramatically; simply "being able to code" no longer constitutes a core competitive advantage. But good engineers are more than just coders—they understand systems, understand constraints, and know which solutions will break at scale. That judgment remains scarce.
Designers face a similar situation. AI can generate a decent-looking UI in seconds, but the gap between "looks good" and "feels right to use" still requires a human to bridge. Real design ability isn't about drawing; it's about understanding what users need in specific scenarios.
The most awkward position may be that of middle managers. Much of their value previously came from "coordinating resources"—allocating limited development resources to the most important projects. Now that development resources are no longer constrained, how valuable is that coordinator role?
The Bottleneck Is Shifting
In the value chain, AI has amplified certain links by tens or hundreds of times; the links that weren't amplified become the new bottlenecks.
Previously, when building products, development took the lion's share; everyone was waiting on dev. Now that development has been compressed, other parts are getting stuck. Do we have good enough ideas? How do we get the product into users' hands? Previously insignificant steps have become the slowest parts of the entire chain. No matter how fast the other links are, everyone has to wait here.
From our own experience, the slowest part of building a feature is no longer development; it's "figuring out what to actually build" and "getting what we built in front of users." Product definition and distribution—these two ends are the bottlenecks. Development in the middle has become the fastest part, so fast that we often finish building only to realize we were heading in the wrong direction.
Product Managers Have Changed
"Everyone is a product manager" is no longer just a slogan.
The product manager of the past was a coordinator who managed processes and helped others realize their ideas. What's needed now is a different capability: the ability to judge what is valuable, to break a broad direction into the smallest testable hypotheses, and to test them one by one—correcting when wrong, building on top when right.
What's scarce about this? It's not methodology. It's will. The willingness to get your hands dirty, direct agents to validate ideas, hit walls, keep adjusting, and keep trying.
When you have this will, AI feels completely different. It's not about "efficiency gains." It's that my ideas can become reality. Previously, you had to persuade people, fight for resources, and wait for scheduling. Now you don't need to convince anyone. Productivity is abundant; what's scarce are people who know where to apply it.
So if you ask me, in this era of productivity surplus, what should you invest in most? It's not learning more tools, not chasing newer models. It's cultivating your ability to ask good questions. Knowing what to validate and where to focus—that is the truly scarce resource.