The AI Coding Tool Landscape (And Why I Just Switched Tools)

I didn't think it would happen.

For months, Anthropic's Claude Code has been my daily driver for agentic coding (what most would call “vibe coding”). I've written about it, recommended it, and built with it constantly.

But this week, I switched to OpenAI's Codex CLI.

Not because I wanted to -- I was pretty happy with Claude Code. But after too many context window issues and rounds of the agent confidently claiming it "fixed" things that still didn't work, I wanted to try something else.

The switch got me thinking about how fragmented the AI coding space has become. There are now dozens of tools, each solving slightly different problems. And most people don't understand the differences.

So today, I want to break down:

• The AI coding tool landscape (and what each category actually does)

• Why I switched from Claude Code to OpenAI’s Codex CLI

• Where AI-assisted coding is headed

—@dharmesh

The AI Coding Tool Landscape

The AI coding space has exploded in the past year.

But most people lump all the AI tools together when they're actually solving very different problems.

Here's how I think about it:

Vibe Coding Apps

Tool ecosystem: Lovable, Replit, Bolt, Vercel v0

These are what I call "vibe coding" tools. You describe what you want in natural language, and they generate an entire application for you.

They're incredible for:

• Prototyping ideas quickly

• Non-developers building simple apps

• Getting from zero to something working in minutes

They're best for net-new projects or quick apps. If you need to work on existing codebases or make production-level apps, you'll hit limitations quickly.

CLI Agentic Coding Tools

Tool ecosystem: Claude Code, Codex CLI, Warp

CLI stands for "Command Line Interface" -- basically, these tools live in your terminal and work alongside your existing development setup.

The key difference from straight vibe coding generation apps: these are for developers actively building, not just for generating apps from scratch. (I fall into this camp of people — a professional developer that uses AI as a dev teammate).

They help you:

• Implement features in existing codebases

• Refactor code

• Debug issues

• Work across multiple files

IDE Integrations

Tool ecosystem: Cursor, GitHub Copilot

These AI tools integrate directly into your code editor (IDE). Instead of switching between your terminal and ChatGPT, the AI lives right where you're already working.

Cursor's Composer mode is particularly impressive -- it can edit multiple files simultaneously while you watch the changes in real-time.

The trade-off with these tools is that you're locked into using their specific editor, which may or may not fit your workflow.

Auto-Debugging Tools

Tool ecosystem: Devin, Cascade (Windsurf)

The newest category. These tools watch your codebase continuously and can:

• Automatically review pull requests

• Suggest fixes for bugs

• Monitor for code quality issues

• Sync changes across branches

This is where things are heading -- agents that don't just respond to prompts but actively monitor and improve your code.

Quick note: Although these categories are distinct, the lines can start to get blurry. You can use Codex CLI to build an app from scratch. You can use Claude Code (the terminal version) inside VS Code. There’s a web-based version of Codex such that you imagine just tagging issues right inside GitHub and having it go review, fix and submit a pull request — just like a developer might.

Why I Switched from Claude Code to Codex CLI

While all the AI coding tools have their place in the market, the category I use the most by far is CLI Agentic Coding Tools.

I recently made the switch from Claude Code to OpenAI’s Codex CLI, and wanted to share my experience since it might resonate with some of you.

Here's what kept happening: I'd give Claude Code a straightforward feature to implement. Nothing crazy -- the kind of thing I know it (and AI coding platforms in general) are actually capable of doing.

It would start working. Make progress. Then hit a brick wall and go down a rabbit hole.

I'd nudge it back on track. It would claim to "fix" the issue, telling me confidently that everything works now. Except it didn't.

Ten-plus rounds of this. Same pattern every time.

The other issue was context window compaction. I found myself spending a decent chunk of my day just waiting for Claude Code to compact its context window so it could continue.

This happened with both Sonnet 4.5 and Opus -- switching models didn't help.

I'd been using Codex CLI occasionally for specific tasks -- it's always been really good at thinking through implementation details and refactorings. But I never made it my default because I was comfortable with Claude Code.

After switching fully to Codex CLI for real work, here's what I noticed: It's gotten noticeably better recently. The quality has jumped.

It actually finishes tasks without getting stuck in loops. When it says something is fixed, it usually is.

I tried Codex when it first launched six months ago while adding features to Agent.ai. It was useful then, but had quirks. Those rough edges are smoothed out now.

I also tried Cursor 2.0 during this process (because why not test everything while I'm at it?). The IDE integration is honestly the best I've seen, and Composer-1 is impressive and fast.

But for my workflow, Codex CLI ended up being the best fit.

What this tells us about agentic coding

The interesting thing about this whole experience isn't really about Claude versus Codex. It's about how quickly things are evolving in this space.

A few months ago, the tools were clearly differentiated. Claude Code was the obvious choice for certain things, Codex for others. Now, the gaps are closing fast, and subtle differences in reliability matter more than headline features.

What I'm watching for:

• Reliability over flashiness -- When an agent tells you something is done but it's not, it’s a dealbreaker. The tool should consistently deliver on what it says.

• Memory and learning -- These agents start fresh every time. The one that figures out how to learn from past projects will have a massive advantage.

• Speed of iteration -- Things are moving so fast that we may see new developments very soon.

My "default tool" might change again next month. That's both exciting and exhausting.

I'll try Claude Code again in a few days to see if the issues were temporary. That's what I love about this space -- by the time you finish writing about a problem, it might already be solved.

But for now, Codex CLI is what's working for me. And in the end, that's all that matters -- not which tool you want to use, but which one actually lets you build.

What tools are you using for agentic coding? I'd love to hear what's working (or not working) for you.

—Dharmesh (@dharmesh)