The Claude Code updates over the past few months have been significant, and if you’re using it daily like I am, you’ve probably noticed things getting noticeably better. Here’s a roundup of the biggest changes.
Agent Teams: Multi-Agent Collaboration
This is the one that caught my attention the most. Claude Code now has a research preview of Agent Teams — you can spin up multiple agents that collaborate on different parts of your codebase simultaneously. Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and you’re in.
Think of it like having three developers working in parallel: one refactoring your API layer, another writing tests, and a third updating documentation. Each agent has its own context and can be configured with custom prompts, tool access, and even different models. They also support memory frontmatter, so agents can maintain persistent knowledge across sessions.
It’s token-intensive, but for large refactors or multi-component features, the time savings are real.
MCP Tool Search: The Context Window Fix We Needed
This was probably the most impactful quality-of-life improvement. Before this update, if you had several MCP servers connected, their tool definitions could eat up 50,000+ tokens before you even typed a prompt. One Docker MCP server alone could consume 125,000 tokens just defining its 135 tools.
The fix is elegant: lazy loading for AI tools. Claude Code now dynamically fetches tool definitions only when needed instead of preloading everything. The results speak for themselves — accuracy on MCP evaluations jumped from 49% to 74% on Opus 4, and from 79.5% to 88.1% on Opus 4.5.
Boris Cherny, Head of Claude Code, put it simply: every user now gets way more context, better instruction following, and the ability to plug in even more tools.
Claude Opus 4.6 with Fast Mode
The underlying model got a major upgrade. Opus 4.6 brings improved coding skills and is available in Claude Code with a new Fast Mode that makes quick operations feel much snappier. Combined with better instruction following and fewer wasted tokens, everyday coding sessions feel noticeably smoother.
For those on the Opus 4.5 model, the improvements are also substantial — state-of-the-art on real-world software engineering benchmarks, 65% fewer tokens on long-horizon tasks, and pricing at $5/$25 per million tokens.
Sub-Agents and Custom Skills
Claude Code now supports sub-agents — independent agents that handle discrete parts of a parent agent’s task. You can restrict which sub-agents can be spawned, configure them with custom prompts and tool access, and they run in parallel with their own context.
The Skills system has also matured. Skills defined in .claude/skills/ load automatically from additional directories. The character budget scales with your context window (2% of context), so larger windows mean more skill descriptions without truncation. Teams can create reusable, task-specific modules and share them across projects.
Claude in Chrome
A beta feature that’s surprisingly useful — Claude Code can now control your browser directly through the Chrome extension. The /chrome command lets you test and debug web applications without leaving your terminal workflow. It’s still early, but for frontend work it reduces context switching significantly.
Other Notable Updates
- Automatic memory: Claude now records and recalls memories as it works
- “Summarize from here”: Partial conversation summarization when you’re running long
- LSP support: Real-time diagnostics for better code accuracy
- Improved Bash tool: No more “Bad substitution” errors with JavaScript template literals in heredocs
- Sandbox improvements: Better permission handling and security
Claude Code has gone from a useful terminal assistant to something that genuinely changes how I approach larger projects. The combination of agent teams, efficient tool management, and the Opus model improvements makes it a serious contender for anyone doing substantial development work.
What features have made the biggest difference in your workflow? I’m particularly curious if anyone has been experimenting with Agent Teams on real projects.