- Published on
Claude Code 2.1.323: Experimental Code Mode and Ultrathink xhigh β The Evolution of AI Coding Tools
"What if, instead of teaching AI how to use tools, you let it write code to call them directly?"
That is exactly the idea behind Claude Code 2.1.323.
Experimental Code Mode: Claude Calls Tools via JavaScript
Previously, Claude Code invoked tools through JSON commands. Starting with version 2.1.323, an opt-in Experimental Code Mode has been added.
In this mode, instead of issuing tool calls directly, Claude writes JavaScript code to invoke tools. In simple terms, Claude has stepped up from "someone who gives instructions" to "someone who writes the code themselves."
Why does this matter? When handling complex tasks that chain multiple tools together, expressing that logic as code gives Claude far more precise control and enables natural looping and branching. Code is simply a better fit than JSON for multi-step automation.
Since this is still experimental, plugin version validation has also been tightened to prevent unintended plugin versions from running.
Ultrathink Now Runs at xhigh Effort
The /ultrathink command in Claude Code puts Claude into a deeper thinking mode. This update changes that command to run at xhigh effort, one level higher than before.
A new per-subagent effort setting has also been added, letting you direct more reasoning resources to specific sub-tasks in complex multi-agent workflows.
This granularity signals that AI tools are maturing β moving from "use it or don''t" toward "how deeply should it think about this."
Auto Mode Changes: Clearer Permission Boundaries
Auto mode no longer falls back to the built-in classifier. Calls that the server cannot verify are now either asked about or denied outright.
Another notable addition: a new check asks for confirmation before a web or shell tool sends the user''s account email address over the network. It is a quiet but important safeguard against personal data leaving the session unnoticed.
MCP Tasks Now Run in the Background
When an MCP server triggers a tool call as a task, Claude Code immediately moves it to the background and reports back when it finishes β no need to wait on-screen for long-running operations.
Opus 4.8 Fast Mode Ends November 30
Important notice: Opus 4.8 Fast Mode is being discontinued on November 30, 2026. After that date, Fast Mode requests on Opus 4.8 will return an error and will not automatically fall back to standard Opus 4.8.
Anthropic recommends migrating to Opus 5.5. Fast Mode has also been newly added on Vertex.
A Note from an EdTech CEO
The most exciting part of this update is the Code Mode. The shift from "command β execute" to "write code β execute" means AI agents can now express far more sophisticated logic.
At the same time, the clearer permission boundaries and the email-transmission check show that Claude Code is maturing not just in power but in trustworthiness and transparency β a critical trait for tools that work deeply inside your workflow.
Tips
- Code Mode must be explicitly enabled in settings. Try it on complex multi-step automation tasks.
- Use Ultrathink selectively β applying xhigh effort to simple tasks will slow responses unnecessarily.
- If you are on Opus 4.8 Fast Mode, plan your migration to Opus 5.5 before November 30.
Sources: