AI Made Coding Faster. Why Does Shipping Still Feel Slow?
The thing that bothers me about working with AI is how easily it goes in a direction that makes sense technically but feels disconnected from the actual product.
It can write the code, add tests, explain the architecture, and still leave me thinking: why are we building all of this?
That's where a lot of the work still sits.
Busy solving the wrong problem
Say you ask an agent to improve a document upload flow.
It could add a queue, build a reusable upload manager, introduce more configuration, and handle a dozen hypothetical file types. Each decision has a reasonable explanation.
Meanwhile, the person using the product just wants to know whether their PDF uploaded and what they can ask about it.
A filename, clear progress, and a useful first question might have solved their problem.
The agent needs direction about which of these things matters. “Improve this” leaves a lot of room for technically convincing work that adds very little value.
The work moves somewhere else
A study by Fiona Chen and James Stratton, covering 718 firms, found that adopting coding agents increased code output. But the gains didn't translate proportionally into completed software. Review became a bottleneck.
The study uses data through March 2026, so it doesn't settle what today's tools can do. But the distinction is useful: generating more code creates more work to understand, check, and maintain.
Someone still has to ask whether the change belongs in the product. Whether it preserves an existing workflow. Whether the additional complexity is worth keeping.
That takes attention, even when writing the implementation took minutes.
“Make it better” needs a definition
I think useful direction starts with describing what a person should be able to do.
For the upload example:
“Someone should be able to upload a PDF, see when it's ready, and ask their first question without needing instructions.”
That gives us something concrete to check. We can open the page and try it.
It also helps to state the constraints. Keep the existing storage. Support PDFs for now. Ask before adding a dependency. Handle a failed upload clearly.
These details narrow the work enough for the agent to make useful decisions without needing approval for every line.
Knowing when to stop
The harder question is often whether another improvement is worth making.
An agent can keep finding things to change. More abstraction, another setting, a cleaner implementation. Each pass can look productive while making the result harder to follow.
I want AI to help me move faster through the work that matters. That means giving it a clear problem, checking the result against that problem, and being willing to say: this is enough.
Writing code has become much easier. Deciding what deserves to exist still needs care.