I went from doing everything in Cursor to running multiple AI agents across projects. Here's exactly what I use, when, and the budget-friendly alternatives.
I made a video about this recently but figured I'd write it up properly for the dev.to crowd since there's more nuance I wanted to get into.
About 6 months ago, my entire AI coding workflow was just Cursor. Claude models, composer, done. It handled UI, debugging, features — everything lived in one place.
That setup already feels ancient.
There's a pretty clear progression I've noticed most devs going through right now:
IDE-based AI (Cursor) → Terminal agents (Claude Code) → Agent orchestration (Codex desktop, Claude Code desktop)
I've been through all three. Here's where I've landed.
Cursor — still use it, just differently now
Not going to pretend I've abandoned Cursor. It's still great for quick edits and when I want AI assistance without leaving my editor.
The one change: I default to Composer 2 for basically everything. I know the launch was controversial but honestly — it's nearly Opus-level for 99% of daily tasks and way cheaper. If you're still manually picking models per prompt, just switch to Composer 2 as your default and forget about it.
But Cursor alone isn't enough anymore. It's one layer of the stack now, not the whole thing.
Claude Code — where I do all my UI work
This was the bigger mental shift. Building full apps from your terminal without opening a single file in your editor is wild when you first experience it.
Where it really clicks for me: UI and frontend work.
Once you connect MCPs and agent skills — particularly Figma MCP — you can go from Figma design → pixel-perfect UI directly through Claude Code. That pipeline alone changed how I approach frontend tasks.
If you haven't set up MCPs for Claude Code yet, seriously do it. It's a multiplier.
Agent orchestration — where I think everything is heading
This is the part that feels newest and most underrated.
Tools like the Claude Code desktop app and Codex desktop app let you run and manage multiple agents across multiple projects at the same time. Instead of one terminal, one agent, one task — you spin up several and keep tabs on all of them.
My split:
Codex → backend, architecture, system design stuff
Claude Code (desktop app) → UI and frontend work
Running Claude Code through the desktop app instead of raw terminal means I can manage multiple instances and actually see what each agent is doing. That visibility matters when you're juggling things.
This separation has been my biggest productivity unlock this year. Stop forcing one tool to do everything.
The budget stack (if subscriptions are killing you)
All of the above is expensive. Here's what I'd do on a budget:
Open Code instead of Claude Code. Runs in terminal, similar workflow, but gives you access to non-Anthropic models that are generally cheaper.
Codex subscription if you're going to spend money on one thing. Best output-per-dollar ratio right now IMO. Cursor + Claude Code costs stack up fast.
Pair those two with something like Windsurf and you're covering most use cases at a fraction of the cost.
The thing nobody talks about
Half of being productive with AI tools is just knowing what exists. The space moves weekly. New MCPs, new agent frameworks, new models, new workflows that make last week's setup feel inefficient.
I got tired of checking Twitter, Reddit, HN, and a dozen newsletters every morning trying to keep up. So I built Trace — it pulls from 100+ sources daily and gives you a curated, summarized view of what's actually happening in AI and tech. No infinite scroll, just the stuff that matters. Might save you some time if you're in the same boat.
TL;DR
The tools will keep changing every few months. The habit of matching the right tool to the right task is what actually sticks.
Would love to hear what setups other people are running — especially if you've found good orchestration workflows. Drop a comment.
Top comments (0)