---
title: "Blog"
description: "Read about the background of our monitoring solution, new features and case studies"
canonical: "https://www.rumvision.com/blog/anthropic-made-claude-3x-faster-the-interesting-part-rum-ai/"
---

Breadcrumbs: [Home](https://www.rumvision.com/?format=md) > Blog

# Anthropic made Claude 3x faster. The interesting part? RUM + AI

Anthropic recently shared how their engineering team made the core Claude experience roughly 3x faster in just two weeks.

Very impressive. But what caught our attention wasn't only the 3x.

It was **how they got there**.

Their engineers gave Claude access to increasingly granular performance measurements, used it to investigate bottlenecks, shipped changes, and measured again.

Or, as Anthropic puts it:

> Once Claude can measure something, it can make it faster.

And that might be the best explanation yet of why we are so excited about combining AI with Real User Monitoring.

## A green Core Web Vital doesn't mean every user is happy

### Their CLS was green. Their sidebar was still jumping.

One of my favorite examples from Anthropic's case is CLS.

Their CLS measurement was already 0.008. Very comfortably in the green. Yet users were still seeing a sidebar jump around.

To find out why, they needed more than the score. They needed to know **which DOM elements shifted and when**. And this is exactly where granular RUM becomes so important.

### A score tells you there is a problem. RUM helps you find it.

A Core Web Vitals score can tell you *that* users have a problem. Granular RUM helps you investigate **who is experiencing it, where it happens and what is causing it**.

We see the same with INP. A poor score tells you where to look. RUM can help you find the interaction, element or script behind it.

## AI knows web performance. It doesn't know your users.

AI knows a lot about web performance. But it doesn't automatically know what happened on **your website, for your real users, yesterday after your latest release**.

Without access to that data, it can suggest what to investigate. But it is still missing the evidence.

### Give AI the evidence and the questions change

Give it detailed RUM data and things get much more interesting.

Now an AI agent can investigate questions such as:

- Why did INP get worse after Tuesday's release?
- Which elements or scripts are behind the regression?
- Did the fix we shipped actually improve the experience for real users?
 
That's why Anthropic's story resonates so much with what we've been building at RUMvision.

## RUM collects the evidence. MCP lets your AI investigate it.

Traditionally, you would investigate RUM data yourself: open dashboards, select segments, compare periods and inspect elements and scripts.

### From dashboards to questions

With MCP, that interface changes. Your AI can access RUMvision data while investigating a question and retrieve the relevant evidence itself.

 [Watch video from www.rumvision.com](https://www.rumvision.com/file/upload/img/video/mcp-flow-orbits.webm)In short:

1. You ask a question;
2. and your AI sends it to the RUMvision MCP;
3. The MCP retrieves the relevant data and returns it together with a tailored guide written by RUMvision;
4. This gives your AI the data and context it needs to make sense of what happened.
 
### You need both RUM and a way for AI to access it

The combination needs two things: **granular RUM data**, and a way for your AI to access it.

Without RUM, your AI is missing the real-world evidence. Without MCP, you're still doing most of the digging yourself.

Put them together and you get a workflow that looks remarkably similar to the one Anthropic describes:

Measure  Investigate  Change  Measure

## This is the part we're most excited about

### This isn't a RUMvision case study

Anthropic obviously has its own engineering teams, infrastructure and instrumentation. They weren't using RUMvision, and their setup isn't something we're trying to claim as a RUMvision case study.

But it is a very good showcase of the direction we believe performance engineering is heading.

### AI alongside performance engineers, not instead of them

Not AI replacing performance engineers. Not another AI chatbot giving generic advice based on a Lighthouse report.

But AI working alongside developers, with direct access to detailed evidence from real users.

The better the measurements become, the better the questions an agent can answer. And apparently, once it can measure something, it might just be able to make it faster.

Want to read their full case study, or start doing the same with your own RUM data?

[Anthropic's case study](https://claude.dev/blog/how-we-made-claude-ai-faster/) [about RUMvision MCP](https://www.rumvision.com/features/web-performance-mcp/?format=md)