I started using AI coding agents seriously about six months ago. Before that, I was skeptical — I'd tried ChatGPT for code and got back plausible-looking nonsense that didn't compile. Then I tried the newer generation of tools — agent-based ones that can actually read files, run commands, and iterate. The difference is night and day.
The good
The biggest win is speed. Things that used to take a full evening — wiring up a backend endpoint, setting up a project — now take minutes. It's like having a junior developer who works 100x faster.
Where it really shines is exploration. Want to try three different approaches? Just ask. Each one takes 30 seconds instead of 30 minutes. You learn faster because you can compare working implementations.
The bad
The number one problem: AI generates code that looks right but has subtle bugs. Off-by-one errors, wrong API calls, edge cases it didn't consider. Every output needs review.
Context management is another issue. Agents work great on small, focused tasks. Give them a large codebase and tell them to "add a feature" and they'll often break unrelated things or reinvent wheels you already built.
What I've learned
- Be specific. "Add a login button that redirects to /dashboard after auth" works. "Add a login button" is a coin flip.
- Review everything. AI is great at first drafts, terrible at final ones.
- Use it for boring stuff. Boilerplate, tests, configs — tasks you know how to do but don't want to. That's where AI saves the most time.
- Keep files small. AI works better when it can read relevant context without hitting limits.
It's not going to replace developers. But it's absolutely going to change what being a developer looks like.