The Two-Hour App Is a Two-Week Problem
"I gave it the specs and had a working app in two hours" sounds impressive, until you look at what happened in those two hours.
The internet is full of posts like this one:
"I gave it the specs, and in two hours I had a working app."
It sounds impressive. It isn't, once you look at what actually happened in those two hours.
What the two hours really contained
- The agent read the same files fifteen times.
- It made a change, broke something, and "fixed" it by working around it.
- It hallucinated a dependency that doesn't exist.
- It spent forty minutes going in circles on a single error.
- A person watched, anxiously, nudging it back on track.
- The "working app" has no tests, questionable architecture, and nobody fully understands what was built.
That's not a feature. That's a junior developer left alone for two hours with no supervision.
The follow-up story nobody posts
Nobody posts the next part: three days of debugging, a rewrite of the parts nobody understood, and the architecture review that found it couldn't scale. The visible cost was the two hours. The invisible cost is the rework, the drift, and the fact that nothing was learned. The next session starts from zero.
Velocity comes from not wasting time
Autonomous agents are fast at typing. They're slow at thinking, and they do that thinking on your clock, in your budget, without your institutional context.
The better model isn't "AI that codes." It's a senior engineer running a capable contractor:
- Draw on what's been tried before.
- Know which parts of the system are fragile.
- Write a precise brief that leaves no room for guessing.
- Review the result before it ships.
- Write down what was learned, so the next job starts ahead.
When the hard thinking is done first, with full knowledge of the business, the actual code-writing takes seconds. The AI becomes a very capable typist, and the work compounds.
The compounding gap
| Session 1 | Session 10 | Session 50 | |
|---|---|---|---|
| Stateless agent | Learns the system | Starts from zero again | Still starts from zero |
| Knowledge-first approach | Learns the system | Knows prior decisions and failures | Institutional memory rivaling a two-year team member |
Every session that leaves nothing behind is a session you'll pay for again. The goal isn't a faster first draft. It's never having to start over.