The question I get most from junior developers in Colombo right now is some version of "should I even bother learning to code, given AI coding agents?" So it caught my attention when the founder of the company behind one of the most capable agents said, plainly, that his tool is not meant to replace us.
Scott Wu, CEO of Cognition (the company that makes the AI coding agent Devin), told TechCrunch that Devin isn't designed to supplant human programmers. Coming from the person who profits most from the opposite being true, that's worth sitting with.
🔍 Why this matters more than the usual "AI won't take your job" line
We've all read the comforting takes. This one is different because of who said it. Wu is a famously strong competitive programmer running a company whose entire value depends on Devin being good. He has every commercial incentive to claim the agent replaces headcount. He didn't.
That tells me the honest internal view at a frontier lab is closer to augmentation than substitution — at least for now. The agent is a force multiplier on a human who knows what they want. It is not the human.
Key takeaway: When the people building the most aggressive automation tell you the tool needs a skilled operator, believe them over the hype cycle that's trying to sell you a course.
For a Sri Lankan engineer billing in USD or a student deciding what to study, the practical reading is this: the skill that pays is no longer "can you write the loop." It's "can you decide what should be built, judge whether the output is correct, and own the result."
📊 What the agent is good at vs. what it can't carry
After using these tools daily, here's roughly how the work splits. This isn't from the article — it's my own working breakdown, so treat it as opinion, not gospel.
| Task | Agent handles it well? | Who still owns it |
|---|---|---|
| Boilerplate, scaffolding, config | Yes, mostly | You skim and approve |
| Writing tests for known behaviour | Often | You decide what to test |
| Translating a clear spec to code | Often | You wrote the spec |
| Deciding what to build | No | You |
| Judging if the output is actually correct | No | You |
| Owning the bug at 2 a.m. before a release | No | You |
The pattern is consistent: the agent compresses the typing, not the thinking. The bottleneck moves up the stack to judgment, and judgment is exactly the thing you can't fake when production breaks.
💡 What this changes for how juniors should learn
Here's where I'd push back on the doom. The risk isn't that agents take junior jobs. The risk is that juniors use agents as a crutch and never build the judgment that makes the agent useful in the first place.
If you let the tool write everything while you're still learning, you get fluent-looking code you can't debug. That's the trap. My advice for anyone learning right now:
- Write it yourself first, then compare. Solve the problem by hand, then ask the agent how it would do it. The gap between your version and its version is your lesson.
- Read every line the agent produces. If you can't explain a function it wrote, you don't own that code — you're just hosting it.
- Break things on purpose. Run the generated code, feed it bad input, see how it fails. Understanding failure modes is the senior skill.
- Practise the fundamentals where it's free. You don't need a paid sandbox to drill loops, recursion, and data structures.
That last point is partly why I built a few free, no-signup playgrounds. If you're learning Python and want to test a snippet the agent gave you without trusting it blindly, you can paste it into our online Python compiler and watch it run in your browser. Same for C and C++ if that's your university track. The point isn't the tool — it's the habit of verifying instead of assuming.
🌐 The free-tier angle for small Sri Lankan teams
For a two-person startup in Galle or a freelancer working solo, the augmentation framing is genuinely good news. An agent that drafts the boring 60% lets a tiny team punch above its headcount without hiring you can't afford.
But there's a discipline tax. A few things I'd hold firm on:
- Never ship code you don't understand. The agent doesn't get fired; you do.
- Keep a human in the loop on anything touching money, auth, or user data. Wu's whole point is that the human stays responsible.
- Don't pay for capability you're not using yet. Free tiers and open-source models are enough to learn the workflow before you commit budget.
The teams that win with these tools won't be the ones that fire their developers. They'll be the ones whose developers got better at the parts a model can't do.
🛠️ What this means for you
If you're a student: keep learning to code, and learn it properly. The fundamentals are now more valuable, not less, because they're what lets you steer the tool instead of being steered by it. The job title might shift from "person who writes code" to "person who decides what code is right," and that's a more interesting job anyway.
If you're a working engineer or a small team: adopt the agent for the typing, keep the judgment in-house, and never outsource ownership. The founder of one of the best agents on the market just told you the human is still the point. Take the free win, skip the panic, and go build something only you would have thought to build.
Top comments (0)