AI works more autonomously, and you set the boundaries

This week's stories paint a clear picture: AI tools are getting better at acting on their own, while users get more choices about how much they want to leave to the machine and how much they should control themselves.

This is most clearly seen in the new agent features. Musk's Grok Bot can now save repeated tasks as fixed "routines," so the AI carries them out itself next time. The same direction is seen at GitHub, where Copilot now lets more users decide how thoroughly the tool should check their code, and the feature is available on all subscriptions.

Cheap power, but also self-reliant solutions

Another thread is about saving money. DeepSeek's cheap language models can now be used directly inside Anthropic's coding tool Claude Code, which according to the story can cut up to 98 percent of the bill. Anthropic has at the same time cleaned up its own house: an updated tool automatically removes outdated instructions from prompts, so companies both save money and get better answers from the new model Opus 5.5.

In parallel, interest is growing in running AI models yourself instead of using a cloud service. A new guide shows in 13 steps how to install Alibaba's free image model Qwen-Image-2.1 locally, including in the free tool ComfyUI. It gives more control, but the model is still only for research, not commercial use.

Trust is the hard part

But the more AI can do on its own, the more important the question of what you can trust becomes. A fake AI image of the US president was shared millions of times before it turned out to be fabricated, and the case shows how hard it actually is to expose such images. Conversely, the creator of Ruby on Rails, David Heinemeier Hansson, is going the opposite way. He has not written a line of code by hand since March and now calls AI the biggest thing to happen in computing.

The pattern in this week's stories is thus not that AI does everything for you, but that it does more if you set the boundaries yourself, choose the right tools and keep an eye on the result. In the coming week, it may be a good idea to test a single new feature, such as a routine in an agent or a cheaper model in a coding tool, rather than replacing the entire workflow at once. And when something seems too good or too convincing to be true, like the image of the president, it is worth double-checking the source before sharing it further.