A developer podcast host recently said they only use AI for autocomplete. This shocked me.

That’s two generations behind today’s state of the art. This is how the field is evolving:

AI Programming Paradigms at a Glance
  1. Super autocomplete: AI predicts code, not just keywords.
  2. Conversational Coding: Chat with your IDE, direct the AI, iterate together.
  3. Agentic Coding: AI acts independently - runs commands, checks work, iterates.
  4. Simul Agentic? : Multiple sub-agents, parallel workflows, working together.

I. Super autocomplete

This used to be one of the biggest selling points of certain IDEs like JetBrains & Visual Studio1.

The first wave of AI changed this. IDEs now predict entire classes and logic blocks, not just keywords. They excel because they feed surrounding code context to the LLM for relevant suggestions.

GitHub deserves credit for kicking off the revolution with their Copilot offering. Here’s an early YouTube video of mine from June 20222 showing it in action.

Autocomplete keeps advancing. Cursor (today’s most popular AI IDE) has “Tab”. JetBrains, the pre-AI autocomplete champion, is building Mellum (an LLM built for code completion).

This paradigm is alive and well, but it’s become table stakes. Most developers use it, but it’s far from the frontier.

II – Conversational Coding

When ChatGPT took the world by storm, a new paradigm emerged. Conversational coding, where you chat with the AI and pair program together.

Unlike the autocomplete era, where you trust the AI’s suggestions, here you direct the AI, give it context, and nudge it toward better solutions. This is arguably what most devs use today and envision when they hear “AI programming”.

It feels magical, productivity jumps are real, and it’s hard to think of every going backwards. Cursor leads this charge: chat with your IDE and have it make code changes on the fly.

The challenge with conversational coding though is IDE lock-in. Android developers love Android Studio, iOS developers love Xcode. Asking them to switch to Cursor is a big ask. There are workarounds: Firebender offers Cursor-like functionality to existing IDEs, JetBrains has Junie (though they’re late to the party), and Xcode developers… well, they wait for Apple.

This constraint is why the next paradigm is starting to shine.

III – Agentic Coding

This is the bleeding edge.

Agentic coding works anywhere you can run commands. Pop open a terminal, fire up Claude Code, and let it work independently. The AI comes up with a plan, confirms the plan, makes changes, runs tests, fixes errors—all through your existing CLI tools.

Claude dominates here. The TUI feels retro but works brilliantly for terminal-heavy development. You give the agent a task. It runs commands, checks results, iterates. The feedback loop is tight: ask, execute, verify, repeat.

The space is exploding. Google’s Gemini CLI runs on Gemini 2.5 Pro with aggressive pricing3. Cursor too has a similar take that they call background agent mode. There’s also agentic tools outside of the terminal like Jules (Google) and Codex (OpenAI) that run cloud agents with GUIs. But Claude nails the terminal workflow imho.

The core idea for agentic coding is the same: let the AI act on its own, not just suggest.

So what’s next?

IV – Simul Agentic Coding

Chess prodigy playing simul chess with robot agents

My guess: we’ll move from mono agentic coding to multi sub-agent workflows.

Claude can already do this today4. Think of a chess master playing a (simul)taneous exhibition match: multiple agents building features in parallel, coordinating to avoid merge hell, and you at the center, reviewing each PR in its own tab.

The field evolves at breakneck speed. By the time you read this, this might feel like a mainstay. Revolutionary becomes quaint in two years. Today’s cutting edge is tomorrow’s baseline.

What About Vibe Coding?

Ah, the million-dollar question. Where does vibe coding fit?

Andrej Karpathy (AI Programming Yoda) coined this term. But vibe coding isn’t a paradigm, it’s a style. It’s about trusting the AI to make big decisions with minimal guidance. You can apply it to any paradigm above, but the vibe gets strong in collaborative and agentic modes.

If you’re a programmer wondering where to invest today: don’t get too attached to one specific tool. Try them all liberally. Start getting conversant. It’s too early to say which tool or paradigm will dominate, but AI programming is inevitable and will probably just be what we call programming.


  1. not that one. see intellisense↩︎

  2. Over 2.5 years ago! This is what I mean by “2 generations behind” (in AI development time, each year seems to bring a fundamental shift in how we interact with these tools). ↩︎

  3. Google is offering 60 requests per minute and 1,000 requests per day completely free. For context, Claude’s Pro plan at $20/mo gives you roughly 10–40 prompts every 5 hours. ↩︎

  4. instruct it specifically to “use sub-agents for this workflow”. ↩︎