PN: 2483-AX9
DO NOT REMOVE DURING OPERATION
SN: 002194-C
MAT: AL6061-T6
信頼
Book a call

AI coding tools made writing code cheap. The rest got expensive

If you judge an AI coding tool, or a team using one, by how fast code shows up on the screen, you are measuring the part that got cheap. AI coding tools made generating code close to free. The cost of building a product did not disappear. It moved to the work around the code: deciding what to build, reviewing what came back, and keeping it alive in production.

GitHub made that point itself last week. In a guide to what it calls agentic engineering, the subtitle is blunt: "Writing code got faster. The work around it didn't." The company says it saw more than 1.2 billion commits in a single week in October, over 400% more than a year earlier. More code is landing than ever. The open question is what it costs to turn that code into something you can ship.

The same guide cites a Faros AI study with the number that matters more. When teams moved from low to high AI adoption, their pull requests spent 441% longer in review. The code arrives faster and then waits, because a person still has to decide whether it does what they meant.

Why AI coding tools shift the cost, they do not remove it

Writing a function was never the hard part of software. The hard parts were understanding the problem, fitting a change into a system that already works, and being sure it will not fall over at 2am. AI coding tools are strong at the first and no help with the rest unless a person drives them.

A prototype is genuinely cheap now. A founder can get a first version out with AI tools, and that is a real shift. To be super transparent, I tell founders on calls that if they are better off building that first version themselves with AI, I will say so. The gap shows up later, when the first version meets real users and real data.

What actually gets expensive

Three things, and none of them are typing.

Review is the first. Someone has to read the change and ask whether it does what you intended and what could go wrong. More volume means more reading, which is what that 441% is showing.

Integration is the second. A function that works on its own still has to fit a codebase with its own patterns, auth, and data. This is where AI output most often looks right and is wrong.

Production is the third. The code has to survive real conditions: bad networks, low storage, users doing things you never planned for. This is the part your users actually feel.

We saw a clean version of this with Tapnote, a study app used by over 100,000 students. Their in-app lecture recorder worked in testing and still generated a steady stream of support tickets. Students were losing recordings, sometimes hours of class, to phone calls, low storage, or the app being killed, because the recorder held the whole session in memory and only saved when they hit stop. Fixing it meant rewriting the recorder to flush to disk every 15 seconds, routing native errors into Sentry so the next failure shows up early, and bringing iOS storage checks to parity with Android. That is unglamorous work, and it is what decides whether an app is worth trusting.

, an AI learning platform, had the same problem from the other side. The code was fine, but one shared environment and manual deploys meant every release moved together and engineers waited on each other to ship. We spent four months moving it onto managed AWS, defining separate dev, release candidate, and production environments in Terraform, and putting every change through a pipeline with integration tests and security scanning. Daily cloud cost dropped about 30%. Getting a change safely to production is its own engineering problem, and no AI coding tool touches it for you.

What to measure in an AI coding tool

Stop counting lines and pull requests. They were never the goal, and now they are easy to inflate. Better questions:

  • How long does a change take to go from written to reviewed to merged? That cycle time is the real speed.
  • Who reads the change, and do they understand the system well enough to catch what is wrong?
  • Can you rebuild your environments and roll back when something breaks?

These were always the real work. AI coding tools just turned them into the obvious bottleneck.

What still needs a person

The decisions. Which approach to take. Whether a change does what you meant. Whether to ship it or roll it back. GitHub's own guide lands in the same place: agents can help across the whole workflow, and the judgment stays with people.

Our approach does not change because the tooling got faster. Every engineer who touches your product is senior, we push back when something does not make sense, and we over-document so nothing is ambiguous. A faster way to produce code raises the value of people who can tell good code from code that only looks good.

FAQ

Do AI coding tools replace developers?

No. They make writing code faster and leave the expensive work with people: deciding what to build, reviewing it, integrating it, and running it in production. On a small prototype one person can cover all of that. On a product with real users, it is most of the job.

What should a founder look for in an AI coding tool?

Look past the demo. The tools are all fast at generating code. What separates them in practice is how much they help with review, testing, and the path to production, and how easily a senior engineer can understand and correct what they produce.

If you have an AI-built prototype and you are not sure what it will take to make it real, that is what we do. Book a call and you will get an honest read.

Book
a call

連絡先
If this is the kind of problem you are working through, we can talk it over.