Anthropic pushed a batch of desktop features to Claude Code on February 20 that, taken together, turn the coding assistant into something closer to an autonomous development loop. The tool can now start dev servers, preview running web apps inside the desktop interface, review local diffs before you push, and monitor GitHub pull requests in the background, automatically fixing CI failures and merging when tests pass.
Two days before the update shipped, Claude Code creator Boris Cherny told Y Combinator's Lightcone podcast that the software engineering title would "start to go away" in 2026. The timing feels deliberate, or at least convenient.
The preview loop
The headliner is the embedded app preview. According to Anthropic's blog post, Claude Code on desktop can now launch dev servers and display running applications directly in the interface. It reads console logs, inspects the DOM, clicks elements, fills forms, takes screenshots, and iterates on issues it finds. You can point at a visual element in the preview and tell Claude what's wrong. This is the part that will impress people on Twitter. Whether it handles a messy webpack config or a Rails app with twelve environment variables is another question entirely, and Anthropic's announcement doesn't address edge cases.
The desktop documentation adds some detail: Claude auto-detects your dev server setup and stores the configuration in .claude/launch.json. You can persist cookies and local storage across server restarts. If your project uses a non-standard dev command, you'll need to edit that config file manually. Reasonable enough.
Code review and the PR question
A new "Review code" button lets Claude examine local diffs and leave inline comments in the desktop diff view. Bugs, suggestions, potential issues. You can then ask Claude to address its own comments, which creates a tight loop where the AI is essentially reviewing and fixing its own output. I'm not sure how to feel about that. On one hand, catching an off-by-one error before it hits CI is useful. On the other, an AI reviewing AI-generated code has an obvious blind spot problem.
The PR monitoring is where things get genuinely interesting. For GitHub projects, Claude Code tracks pull request status using the GitHub CLI under the hood. Enable auto-fix and it attempts to resolve CI failures on its own. Enable auto-merge and PRs land as soon as checks pass. The intended workflow: open a PR, move on to your next task, and let Claude babysit the first one. By the time you circle back, either the PR merged or Claude is stuck on something it can't solve.
Sessions that follow you around
Running /desktop in the CLI now brings your full session context into the desktop app. A "Continue with Claude Code on the web" button pushes local sessions to the cloud for mobile pickup. Anthropic clearly wants Claude Code to feel like a persistent collaborator rather than a tool you open and close. The cross-platform sync works across CLI, desktop, web, and mobile.
And then there's worktree support. Boris Cherny announced on February 21 that Claude Code's CLI now has built-in Git worktree support (the desktop app already had it). Each agent gets its own isolated working directory and branch. You can run claude --worktree feature-auth in one terminal and claude --worktree bugfix-123 in another, and they won't step on each other. Subagents can use worktree isolation too.
This matters because it's the infrastructure for parallel AI agents on a single repo. Run five Claude instances on five branches simultaneously. That sounds great until you think about merge conflicts, and I'm curious how well it handles repos with expensive build steps or complex dependency trees.
"Coding is practically solved for me"
Back to Cherny's podcast interview. He told Y Combinator that coding is "practically solved" for him and predicted it would be for everyone by the end of 2026. He said engineers on his team are becoming generalists, with product managers, designers, and finance people all writing code. The title "software engineer" becomes "maybe builder, maybe product manager, maybe we'll keep the title as a vestigial thing."
That's a bold claim from the person who built the product. Business subscriptions to Claude Code have quadrupled since the start of 2026, according to Anthropic, and enterprise use now represents over half of all Claude Code revenue. So the growth numbers are real. But "coding is solved" is doing a lot of work in that sentence. Solved for building CRUD apps with an AI assistant watching over your shoulder? Plausible. Solved for debugging race conditions in distributed systems, or optimizing a hot path in a game engine? I'd want to see it.
Some developers have started flagging the less glamorous side. A software engineer told Business Insider that AI tools have simultaneously made them more productive and more overworked, coining the phrase "AI fatigue." The update ships three days after Anthropic released Claude Sonnet 4.6 on February 17 with improved coding capabilities, so the company is clearly on a shipping tear.
What's actually new here
Strip away the hype and three things matter. First, the preview loop closes a gap that every AI coding tool has: the model edits code but can't confirm the result matches intent. Now it can, at least for web apps. Second, auto-fix plus auto-merge on PRs turns Claude Code into a background process for your CI pipeline, which is a genuine workflow change if it works reliably. Third, worktree isolation makes parallel AI agents practical rather than theoretical.
The update is available now through Claude Code desktop. Whether it delivers on the promise of autonomous development or just adds a slicker UI to the same limitations, well. That depends on what you're building and how much you trust the loop.
