The last 20% decide whether you have a product
Vibe-coding gets you to 80% in minutes. But the last 20% are the whole job — and that's exactly what an agent can't (yet) do alone.
by
Coding agents make the first 80% of a feature nearly free — boilerplate, CRUD, and scaffolding arrive in minutes. The last 20% decide whether that becomes a product or technical debt: edge cases, behaviour under load, testability, and an architecture someone actually understands. That 20% was always the real job; AI just made it more visible. The developer's value shifts from producing code to owning it.
I've spent the last few months working intensively with coding agents — Claude Code, Mistral, ChatGPT. And I'll admit it openly: they do the boring work faster, and often better, than I ever could. Boilerplate, CRUD, glue code, the first scaffold of a feature — minutes instead of hours.
The first reaction is euphoria. The second, if you're honest, is fear: what am I still needed for?
The answer is hiding in a single number.
80% just became free
An agent takes you surprisingly far. One prompt, one run, and you have something that works. Demo-ready. Screenshot-ready. LinkedIn-ready.
But "works in the demo" and "carries a business" are two different universes. Between them sit the last 20%:
- What happens on the edge case nobody prompted for?
- How does the code behave under load, with real data, in six months?
- Who understands the architecture when something is on fire?
- Is the thing testable, maintainable, extensible — or a house of cards?
That 20% was always the real job. AI didn't abolish it. AI just made the first 80% so cheap that the last 20% is now more visible than ever.
Vibe-coding is real — and ignores 20 years
Vibe-coding works. That's not a sarcastic statement. You describe what you want, you get code. For prototypes, internal tools, throwaway scripts, it's a gift.
The problem isn't the method. The problem is that it pretends the last 20 years of software engineering never happened. Version control, architecture, automated tests, code review, clean failure behaviour — none of that was invented because developers love bureaucracy. It was invented because software without it collapses under success.
An app with ten users forgives everything. An app with ten thousand forgives nothing.
The bridge is the job
I don't believe in "AI replaces developers." I also don't believe in "AI is just hype." Both camps are wrong, because they miss the same thing: the work shifted, it didn't disappear.
The new core skill is no longer typing the 80%. It's owning the 20%:
- Specify, don't prompt. One precise brief beats ten corrective prompts.
- Review as a primary discipline. You read more code than you write — and you have to actually understand it.
- Tests as a contract. The agent can be fast. The tests tell you whether it was also right.
- Architecture in your head. The agent sees the file. You see the system.
This isn't a step back into the past. It's the craft, applied at a new speed.
What this means in practice
If you're a developer: your value moves from producing code to owning code. Get excellent at review. Learn to specify. That survives every model generation.
If you lead a team: allow the AI speed — but don't skip the last 20%. A feature that stands up in an hour and explodes in a quarter was never cheap. It was just prepaid.
The 80% is free now. The 20% is what you still get paid for. That's exactly where I build.
Frequently asked questions
Is vibe-coding good enough to build a real product?
Vibe-coding works — for prototypes, internal tools, and throwaway scripts it's a gift. The problem is that it pretends the last 20 years of software engineering never happened: version control, architecture, automated tests, code review, and clean failure behaviour exist because software without them collapses under success. An app with ten users forgives everything; an app with ten thousand forgives nothing.
What is the last 20 percent in AI-assisted development?
Everything between demo-ready and business-carrying code: the edge case nobody prompted for, behaviour under load with real data in six months, someone who understands the architecture when something is on fire, and whether the code is testable, maintainable, and extensible. AI made the first 80% so cheap that this remaining work is now more visible than ever.
Will AI coding agents replace developers?
Neither 'AI replaces developers' nor 'AI is just hype' holds — the work shifted, it didn't disappear. The core skill is no longer typing the 80% but owning the 20%: specifying instead of prompting, review as a primary discipline, tests as the contract that shows the agent was right, and keeping the system architecture in your head while the agent sees only the file.
What should developers focus on now that AI writes most of the code?
Get excellent at review and learn to specify — one precise brief beats ten corrective prompts, and those skills survive every model generation. Team leads should allow the AI speed without skipping the last 20%: a feature that stands up in an hour and explodes in a quarter was never cheap, just prepaid.