## Coding Tools That Learn Like Humans

This week, AI agents are changing how people write computer code. Think of an AI agent like a super-smart helper robot that knows how to program. Anthropic just released a new version called Claude Sonnet 4.6 that is almost as smart as their best AI model, but costs the same price. This new Claude can understand huge amounts of code at once - imagine if you could give it an entire computer program and ask it to fix something or add something new, and it would understand everything.

OpenAI created something called Codex-Spark that is really, really fast. Regular AI coding tools think carefully through problems, but Codex-Spark is made to be quick. It produces 1,000 new words (called tokens) every second. That means when a programmer types something, Codex-Spark responds almost instantly, like talking to a friend instead of waiting for an answer.

## Big Companies Are Trusting AI to Code

Amazon, one of the biggest technology companies in the United States, made a big decision this week. They told 80% of their computer engineers that they must use AI coding tools at least one time every single week. This is important because it means even huge companies believe AI coding is the future. Some programmers at the company Spotify have not typed a single line of code themselves since December. Instead, AI agents do it for them. Engineers can even push their work live to the internet while they are riding the bus home on their phones!

The companies building these tools use AI to code too. Anthropic now uses AI to write 70 to 90 percent of their own code. Even the OpenAI team said that the Codex tool helped create itself. These are not just dreams about the future - this is happening right now.

## New Tools to Help AI Work on Websites

Google, the famous search company from the United States, released something called WebMCP this week. This tool helps AI agents do things on websites without getting confused. Regular AI has to figure out how a website is built, which is hard and breaks easily. WebMCP is like giving AI a treasure map that shows exactly where things are on a website. Now AI can buy things, make appointments, or fill out forms on websites without making mistakes.

## Problems in Open Source Software

Not everything about AI coding is perfect. Open source software is code that anyone can see and help improve. Many open source projects are having problems because AI coding tools make it too easy for inexperienced people to contribute bad code. The leader of a project called VLC, which is a popular video player, said that many of the new code suggestions are "abysmal," which means really, really bad.

People who work on Blender, a 3D art program, said that AI-suggested code wastes their time. A programmer named Mitchell Hashimoto created a system to only let trusted people contribute code, because AI removed the natural barriers that used to protect open source projects. Even the project called cURL had to stop letting people report security problems because they got so much low-quality AI-generated suggestions that they called it "AI slop".

## The Balance Between Making and Maintaining Code

The real problem is that AI makes creating new code easy, but keeping it clean and working is still hard. Big companies like Meta are happy when their engineers write lots of new features, so they get promotions. But open source developers care more about keeping code clean and stable. AI makes this harder because there is more code to take care of, but the same small number of experienced people has to review everything. One expert said that AI doubled the problem - code is growing faster AND there are not enough experienced people to manage it.