Claude Code's Latest Updates Change How Developers Work

Anthropic just shipped a cluster of updates to Claude Code that fundamentally change how developers can use AI assistance in their workflow. Ray Amjad, who runs a masterclass on the tool, walked through the changes in a recent video—and the implications go beyond just "new features."

The headline feature is git worktree support, which sounds technical until you understand what it solves. Worktrees let you have multiple branches checked out simultaneously in separate directories. For Claude Code, this means you can run parallel AI sessions working on different features without them stepping on each other's files.

"If you had three sessions of cloud code all editing the same file structure without work trees, then one could make a change or move a file around and then the other two sessions would be confused being like, okay, wait, this like suddenly changed," Amjad explains. "But by having it in separate work trees, the file structure stays constant until that thing has been implemented."

The practical use case: You're deep into implementing a complex feature when a colleague needs a quick fix. Instead of stashing your work, you spin up a new Claude Code session in a fresh worktree, handle the urgent change, and your main work remains untouched. Each worktree can have its own sub-agents, its own package installations, its own experiments.

When AI Runs in the Background

The worktree functionality pairs with tmux integration—and this is where things get interesting from a workflow perspective. Claude Code can now run inside tmux sessions, which means it can continue working after you detach from the terminal.

Amjad demonstrates giving Claude Code a research task, detaching from the session, and coming back later to find it completed. For long-running operations—full test suites, comprehensive security scans, extended research tasks—this removes the constraint of keeping a terminal window open and waiting.

The security dimension here matters more than the convenience. Claude Code announced a limited research preview of automated vulnerability scanning that, according to their claims, found over 500 vulnerabilities in open-source codebases—"bugs that had gone undetected for decades despite years of expert review."

Amjad speculates they're using a less restricted model for this feature: "This kind of stuff is a double-edged sword. Like the better the model gets at hacking, the better it gets at identifying security vulnerabilities."

That's the tension with AI security tools. The same capabilities that identify exploitable code paths can be used to exploit them. The question isn't whether to develop these tools—that ship has sailed—but how to deploy them responsibly. Anthropic's approach of a limited research preview suggests they're aware of this.

The Desktop Application Gets Practical

The desktop version of Claude Code gained features that address a specific pain point: verifying that web applications actually work as intended before deployment. The new preview functionality lets Claude Code spin up a local development server, navigate through the application like a user, click buttons, and take screenshots.

"This can be pretty handy for design related work and also getting cloud code to click around and behave like a user for you to make sure things work as intended," Amjad notes.

The element picker feature extends this. You can select any UI element, and Claude Code sees both the visual representation and the underlying code. Change this button to green, adjust this margin, test this form—the AI can make targeted changes without hunting through your stylesheet.

More significant for team workflows: the desktop app now supports SSH connections. You can point Claude Code at a remote server and have it make changes directly. For small side projects, Amjad says he skips git entirely and just has Claude Code work on the server. This violates conventional deployment wisdom, but for throwaway projects or quick fixes, the friction matters more than the best practice.

Memory and Multi-Repo Support

Two features expand what Claude Code can remember and access. Sub-agents can now maintain memory across sessions—at the user, project, or local level. A sub-agent that analyzes logs can remember patterns from previous runs. A code reviewer can track which mistakes it's flagged before.

The multi-repo support lets Claude Code work across multiple repositories simultaneously. The obvious use case is copying implementations between projects, but the more interesting application is comparative analysis—why does this authentication flow work in project A but fail in project B?

What This Actually Means

These updates cluster around a theme: reducing the overhead of context-switching and maintaining multiple development threads. Worktrees, background execution, memory, cross-repo access—they're all mechanisms for keeping more state active without requiring developers to manually manage it.

The risk is the same as with any automation that reduces friction: you can end up with more simultaneous tasks than you can actually oversee. Five worktrees running different experiments sounds productive until you need to evaluate five different implementations and decide which one ships.

The security scanning feature occupies a different category. If Anthropic's claims hold—and 500+ previously undetected vulnerabilities is a bold claim—we're looking at AI that can find classes of bugs that human reviewers miss. Not because humans aren't skilled, but because the pattern-matching required to spot certain vulnerabilities across large codebases exceeds what's practical for manual review.

The counterargument: vulnerability scanners have existed for decades, and they produce false positives and miss real issues. Claude Code's scanning might be better, but "better than existing automated tools" and "finds bugs experts miss" are different claims.

What's clear is that the friction of using AI coding assistance keeps decreasing. Whether that makes developers more effective or just generates more code that needs maintaining—that's the question these tools still need to answer.

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.