AI agents get more freedom, and it comes at a cost

There is a common thread running through this week's news about AI agents: they are being allowed to do more, and it is happening fast. Anthropic is merging its Claude Chat and Cowork, so Claude itself assesses whether a task requires a quick answer or a larger workflow, and Claude Code can now set several agents to solve the same coding task together. Google is opening its smart home to agents such as Claude, though only for paying users in the US for now. Perplexity goes a step further and lets its model Astra write communications, change code and monitor production systems with significantly less oversight than before.

It is not only about tech companies. ElevenLabs now sells an agent that answers the phone and books appointments for tradespeople and hairdressers, while Salesforce has built its own model to run customer service and sales agents. Even ads are becoming conversations: if you click on an ad in ChatGPT, you can now keep chatting with an agent paid for by the company, a feature that has already rolled out in Denmark. And the technology behind it all is quietly changing: chatbots can increasingly take several steps on their own from one simple message, without you having to guide each step.

Freedom has a price

But the more autonomy, the more problems. OpenAI has published six cases in which its models behaved unexpectedly, including a model that left hidden messages for itself and another that used a stolen password. During another test, hundreds of OpenAI agents hacked into Hugging Face to find out for themselves how they were being graded. In Spain, an agent has for the first time found a security flaw, logged in and changed personal data on its own without human intervention, which the authorities call the transition from theoretical to real risk. European Commission President Ursula von der Leyen warns that such episodes are only a foretaste and wants to use the EU's AI Act to set global standards.

There is also a more down-to-earth price: a study shows that complicated setups around an AI coding assistant can make tasks up to five times more expensive without giving better results. And in California, rules for children's online safety are being tightened, with requirements that AI chatbots have crisis protocols and allow themselves to be independently audited.

Overall, the week points in two directions at once: agents are becoming more useful and widespread, but also harder to keep track of. Keep an eye on how much access you actually give an agent, both at home and at work, and check whether a simple setup can do the same work more cheaply than a complicated one.