First came autocomplete. GitHub Copilot, 2021. Your editor guesses the next lines. Developers dismissed it, tried it, couldn't live without it.
Second came chat. Claude, ChatGPT in your workflow. Ask questions, get answers. Paste code, get explanations.
Now comes the third wave: autonomous agents. Systems that don't just suggest—they plan, execute, debug, iterate.
Copilot predicts what you're about to type. Based on your current file, function names, comments.
At best, saves keystrokes. At worst, hallucinates plausible garbage.
You're still driving. AI is just a faster keyboard.
Limitation: inherently local. Sees a few hundred lines. Doesn't understand your architecture.
ChatGPT made AI conversational. Back-and-forth. Explain problems in natural language. Ask follow-ups.
This brought AI into design discussions. "How should I structure this?" became a legitimate question.
Limitation: context windows and integration. Conversations happen outside your IDE. Copy-paste. AI doesn't see your real codebase.
IDE integrations started solving this. Cursor, Windsurf. AI sees your files now. But you're still directing every step.
Agents change the relationship.
You describe a goal: "Add authentication with NextAuth." The agent doesn't just generate code—it plans.
Breaks task into steps. Creates files. Installs packages. Writes config. Handles errors. Tests. Debugs. Iterates.
The AI has agency now. Makes decisions without prompting for each one.
Boring, well-defined tasks:
An agent can produce code that works but is architecturally wrong. It optimizes for the immediate goal without understanding broader constraints.
The value of "writing code" is declining. The value of "knowing what to build" is increasing.
Skills that matter: system design, debugging complex issues, code review, architecture decisions, knowing when AI is subtly wrong.
Skills that matter less: typing speed, memorizing syntax, boilerplate generation.
The developers who thrive will be the ones who learn to ride the wave, not fight it.
— blanho
Everyone defaults to OpenAI. Grab built a 1B model that's faster and cheaper. When should you stop renting?
You type a prompt, an AI writes code. Simple, right? Behind the scenes, it's five systems arguing with each other.
When direct data transfer becomes unwieldy, add a layer of indirection. Netflix learned this the hard way.
