Phase 0: The AI-Native Engineering Mindset · 40 min · Claude Code · Cursor · Cline
Introduction
The Tool Landscape: Durable Categories, Not Brands
A new coding tool launches on X with a viral demo: it builds a working app from a paragraph of English in ninety seconds. Your team lead pings the channel — "should we switch?" Six people reply with opinions based on the demo video. Nobody has actually used it. This happens roughly monthly in 2026, and teams that evaluate every new tool from scratch, by vibes, burn real hours chasing launches that turn out to be repackaged versions of something they already understood.
The fix isn't to ignore new tools — some of them matter. It's to have a mental model that lets you place a brand-new product into a category you already understand within about ten minutes, so you're evaluating "is this a good terminal agent" instead of starting from zero every time.
Four categories, not a hundred brands
Nearly every AI coding tool on the market in 2026 falls into one of four durable categories:
Terminal agents run in your command line, operate directly on your file system, and typically give the model the most autonomy — it can read, write, and execute across your whole repo in one turn. Claude Code, OpenCode, and Cline (which also has an IDE mode) are the current leaders. They're the closest thing to "give the agent a goal and a repo, let it work."
IDE-native tools live inside an editor — usually a VS Code fork — and blend inline autocomplete with agentic multi-file edits, but keep you closer to a traditional review-as-you-go loop. Cursor and Windsurf are the current leaders. The editor context (open tabs, cursor position, project structure) is available to the model without you having to describe it.
Browser app-builders run entirely in a web app: you describe a product, and the tool generates and deploys a full working app, often with its own hosting and preview environment. v0, Bolt, and Lovable are the current leaders. These are optimized for zero-to-prototype speed, often at the cost of you owning the deployment infrastructure directly.
Cloud agents run autonomously on remote infrastructure you don't directly control — you assign a task (often from a ticket or PR description) and the agent works independently, sometimes for hours, then presents a PR. Devin is the best-known example. These trade the tightest human-in-the-loop control for the most autonomous, longest-horizon execution.
Why categories outlive brands
Six months from now, at least one of the specific product names above will have been acquired, rebranded, or overtaken by a competitor you haven't heard of yet. The four categories — terminal, IDE-native, browser builder, cloud agent — are defined by where the tool runs and how much autonomy it has, not by who makes it. Learn to place a new product into a category and you can evaluate it fast, regardless of which company built it.
The reason this course (and the wider "AI-native builder" bootcamp landscape from Maven) deliberately teaches categories instead of doing a deep dive on one specific branded tool is durability. A course anchored entirely to one product goes stale the moment that product's UI changes or a competitor leapfrogs it — and it happens fast enough that plenty of 2024-era "master tool X" courses are now teaching a deprecated interface. Categories are stable because they map to a real constraint (where does execution happen, who's in the loop, how much blast radius does a mistake have), and that constraint doesn't change even as the specific winner in each category does.