TL;DR: Claude Code shipped two rapid-fire versions this week, v2.1.120 and v2.1.121, with the headline features being a new `alwaysLoad` option for MCP server configs, native PowerShell support on Windows, and a `claude plugin prune` command to clear orphaned plugins. These are small version numbers with outsized workflow impact, especially for teams running multi-server agentic setups. One caveat worth knowing before you push to production: there are active GitHub issues on Vertex AI compatibility that could sting Google Cloud shops.

Anthropic has been shipping Claude Code updates at a pace that makes it hard to keep up without a dedicated changelog watcher. This week's back-to-back releases, v2.1.120 and v2.1.121, landed within days of each other and together represent a meaningful step toward what you could call zero-config agentic workflows: setups where the tool gets out of its own way and lets engineers focus on output rather than configuration overhead. Here's what shipped, ranked by impact.

This is the week's most impactful change for teams running complex, multi-server Model Context Protocol (MCP) environments. Previously, tools from an MCP server could be deferred depending on context, meaning your agent might not have the right tools available at the right moment without explicit prompting or configuration fiddling. The new `alwaysLoad` option pins a server's full tool set as permanently available, no deferral, no surprises. For teams with critical MCP servers, things like internal API connectors, proprietary code search tools, or compliance checkers, this eliminates a class of subtle agentic failures where the model simply didn't have access to what it needed. In multi-tool environments, reducing that cognitive and configuration overhead can shave 30-50% off setup friction for developers context-switching between projects. The practical recommendation: identify your two or three mission-critical MCP servers and set `alwaysLoad: true` on them now. Treat it like a pinned dependency. Don't apply it wholesale to every server or you'll bloat the context window unnecessarily.

This one is quietly significant for enterprise shops. Claude Code previously required Git Bash on Windows to function properly, which is a meaningful barrier in organizations where Git Bash isn't standardized or is blocked by IT policy. Native PowerShell support removes that dependency entirely. If you have Windows-heavy development teams, this week's update may be the unblock you've been waiting for before rolling out Claude Code more broadly. The Git Bash requirement was a consistent friction point in enterprise adoption conversations, and removing it matters more than its changelog placement suggests.

Auto-installed plugins accumulate quietly. Over time, orphaned plugins, ones installed automatically during agentic sessions and then never cleaned up, create configuration drift, slow startup times, and introduce debugging headaches. The new `claude plugin prune` command clears them out in a single command. This is hygiene tooling, but hygiene tooling matters at scale. Teams running Claude Code across dozens of developer machines will feel this immediately. Estimate: cleaning up orphaned plugin accumulation can reduce environment setup time by up to 20% for teams that have been running Claude Code for more than a few months without manual cleanup.

Ultrareview is Claude Code's deep code review mode, previously only available interactively. The new `claude ultrareview [target]` subcommand makes it callable non-interactively, which means it can now slot directly into CI/CD pipelines, pre-commit hooks, and automated review workflows. This is the update with the most long-term leverage for engineering leaders. Automated deep review on every PR, without a human having to invoke it manually, changes the economics of code quality at scale. A team of five AI-augmented engineers running automated ultrareview on every commit can now maintain the review coverage that used to require a team three times that size.

A smaller but meaningful capability addition: Skills (reusable behavioral templates in Claude Code) can now dynamically reference the current effort level setting. This lets you build Skills that behave differently at different effort tiers, for example, a refactoring Skill that does a quick pass at low effort and a full architectural analysis at high effort, without maintaining separate Skill definitions. Combined with the new `/skills` type-to-filter search box (also in this release), finding and applying the right Skill in large Skill libraries just got significantly faster.

Before you push v2.1.121 to production on Google Cloud infrastructure, read this. There are two active issues on the Claude Code GitHub tracker that deserve engineering leader attention:

The Vertex AI issue is the more serious of the two for enterprise teams. If your organization routes Claude Code through Google Cloud's Vertex AI backend, hold v2.1.121 in staging until Anthropic ships a fix. Based on typical turnaround on issues at this severity level, budget 2 to 4 weeks before assuming production safety. This isn't a reason to avoid Claude Code. It's a reason to be deliberate about rollout sequencing, which is exactly what good engineering leadership looks like.

Cursor v0.45.x and GitHub Copilot Workspace have both been shipping agentic and multi-platform improvements in parallel. Here's a direct comparison on the features that matter most this week:

The pattern here is clear: Claude Code is pulling ahead on agentic workflow depth, particularly for teams building custom toolchains on top of MCP. Cursor remains the stronger choice for developers who want a polished IDE-native experience without configuration complexity. Copilot Workspace wins on enterprise GitHub integration, full stop. If your team is building or scaling an AI-augmented engineering workflow rather than just adding an AI autocomplete layer, Claude Code's trajectory is the one to bet on.

Concrete actions, ranked by priority:

What Anthropic is building with these rapid Claude Code iterations isn't just a better code completion tool. The combination of `alwaysLoad` MCP configs, effort-aware Skills, and non-interactive review commands points toward something more significant: an agentic development layer that sits above your IDE and orchestrates work across tools, contexts, and team members without requiring constant human steering. The teams that will compound the fastest in the next 12 to 24 months aren't the ones with the most engineers. They're the ones who invest now in understanding how to configure, extend, and orchestrate these agentic layers effectively. That's a skill set, and it's one that separates the engineers worth hiring from the ones who are still treating AI tools as fancy autocomplete. The velocity of Anthropic's shipping cadence, two meaningful releases in a single week, signals that the gap between teams who are actively experimenting and teams who are waiting for the "right time" is widening faster than most engineering leaders realize. The right time is now.