claude-code
Claude Code's New "Code Mode": What It Is and Whether It Matters to Your Business
Anthropic has added an experimental Code Mode to Claude Code that bundles several AI steps into one small script. Here's what it does and what the evidence says so far.
· 6 min read
claude-codeai-agentsai-workflowsautomationanthropicIf you have a developer, an agency or your own team using AI to build and maintain your website or internal tools, there's a good chance Claude Code is somewhere in the mix. Anthropic recently added an experimental feature to it called Code Mode. The name sounds technical, but the idea is simple, and it could affect how quickly (and how cheaply) AI-assisted work gets done.
Here's a plain-English look at what changed, what's still uncertain and what you might ask your tech partner about it.
What Code Mode actually does
Normally, an AI coding assistant works step by step. It searches for a file, reads the result, decides what to do next, edits something, checks the result, and so on. Every one of those steps is a separate back-and-forth with the AI model.
Code Mode changes that pattern. According to the official Claude Code documentation, with Code Mode on, Claude writes a piece of JavaScript that makes the tool calls, and Claude Code runs that code in your session. One script can run several commands in parallel, pass one result into the next step and return only the part of the output the model actually needs.
Think of it like this. Before, the AI would call the warehouse once for every item on an order. Now it sends one complete pick list and gets back a single "all done" reply.
One developer shared a quick demo. He asked Claude to find broken links across a set of documents, and it made 13 tool calls in just 2 model turns and found all 3 broken links. Without Code Mode, that would have been 13 separate round trips.
Where this idea came from
This isn't a random experiment. Back in November 2025, Anthropic's engineering team published a post on code execution with MCP (MCP is the standard way AI agents plug into tools like Google Drive or a CRM). The argument was that AI models are good at writing code, so letting them write small programs to use tools can mean loading only the tools they need and filtering data before it ever reaches the model.
Anthropic also noted that Cloudflare had published similar findings under the name "Code Mode." Claude Code's new feature applies that same thinking to its own built-in tools.
The honest part: does it save money?
This is where you'll want to be a little careful, because some of the numbers going around don't come from where people think they do.
- Anthropic hasn't published a savings figure for Code Mode itself. The Claude Code documentation includes no benchmark or token numbers.
- The big percentages you may have seen come from other work. Anthropic's 2025 post described one example where a Google Drive to Salesforce workflow dropped from about 150,000 tokens to about 2,000. That was a vendor example about a related technique, though, not a test of this feature. A widely shared "37%" figure also comes from a separate Anthropic article on the API.
- Independent tests are mixed. One hands-on test found the fixed "system tools" overhead dropped from 24.6k tokens to 11.1k with Code Mode on. That's a meaningful cut. On the other hand, a critical write-up quotes early users saying it was slower and cost more across several tasks, and that Claude filled up its own memory with logging.
So the fair summary is that it looks promising for multi-step jobs, but the evidence is early and thin. Fewer tokens per request could stretch usage limits on long jobs. Still, nobody should promise you a specific saving yet.
What about safety and control?
This is probably the most important part for a business owner. When an AI is making changes to your website or systems, you want guardrails.
The good news is that every tool call inside the script still goes through the same "deny" and "ask" permission rules and safety hooks your developer has set up. Code Mode doesn't skip them.
There is one catch worth knowing about. In Claude Code's "auto" mode, an automated safety check reviews each piece of code before it runs, but not each individual tool call inside it. The docs also say this check can't see values that are only worked out while the script runs, so they recommend keeping strict rules or hooks for any actions that must stay under control. Some commentators also note that a whole program is harder to review line by line than a single tool call. That's likely part of why Anthropic shipped the feature switched off by default.
The fine print
If your team wants to try it, here's what the documentation and coverage say:
- It arrived in Claude Code version 2.1.323, released November 12, 2026.
- It's experimental and off by default. Anthropic says it may change or remove the feature in future versions.
- It only works in interactive terminal sessions, and it doesn't support Bedrock or gateways.
- Code that Claude waits on stops after 10 minutes. Background code can run for up to two hours, but only in certain permission modes.
- A session started in Code Mode can only be resumed with Code Mode still on.
- Remote Control, Claude in Chrome and computer use don't work while it's on.
- Organizations can block it for everyone through managed settings.
What this means for you
You don't need to turn anything on yourself. But if someone builds or maintains your digital tools with Claude Code, here are a few friendly questions to ask:
- Are you trying Code Mode, and on which kinds of tasks? One guide suggests treating it as something to switch on per session for complex, connected tasks, not as a default for everything while it's still experimental.
- Have you checked your safety rules still hold? Especially for anything touching live sites, customer data or payments.
- Are you measuring the results? Since there are no official numbers, real before-and-after comparisons on your own projects are the only way to know if it's helping.
The bigger trend is worth watching too. AI tools are moving from "one small step at a time" toward writing little programs that do several steps at once. If that pattern keeps holding up, AI-powered workflows for small businesses should keep getting faster and more efficient. For now, though, it's an experiment, so trying it out carefully makes more sense than switching your whole operation over.
Comments