AI agents move into the workplace and become coworkers

An AI agent is a program that doesn't just answer a question but carries out tasks for you. This week's stories show that agents are moving from demo to everyday use. But the picture is not as sharp as the companies' press releases suggest.

From chatbot to coworker

Perplexity has launched Teammates, persistent AI agents that take over ongoing work tasks. They require the expensive Max subscription, though. In Mexico, the cement group Cemex has built an HR agent that around 800 employees can query in Teams. The impact has not been quantified, and the case comes from Microsoft itself. In another Microsoft story, the services firm Huron reports 1,046 active agents within four weeks, but the figure has not been independently verified.

The payroll system Dayforce has also shown a platform with AI agents and a connection to Claude. Parts are in testing, while others are not expected until 2027. Grok can call businesses for you and wait on hold, but only in the US.

Agents learn to use the computer

One clear trend is agents that use a computer the way a person does. OpenAI says GPT-6 Astra can operate programs through the screen. Microsoft has shown Mage-VL-Cowork, a small model that both uses the screen and runs code. According to Anthropic, Claude Code in the cloud now figures out on its own what a project needs and remembers the setup.

DeepSeek is especially active. The open-source program Harness had around 1.54 million daily users, but the figures come from the team itself. The team also describes a Super agent designed to work around the clock. It has been released for free as a research prototype and acts only when an approved person asks it to or a rule allows it.

Memory, control and caveats

When agents are going to work alongside you, they need to remember. Both Living Brain and Membase want to give ChatGPT, Claude and Cursor a single shared memory. These are the makers' own descriptions, however.

At the same time, demand for human oversight is growing. The free tool AiSOC lets agents review security alerts, but humans must approve the actions. A small experiment with 52 students pointed in an interesting direction: AI teammates that disagreed produced slightly better business plans. The result is preliminary. Perhaps a good agent shouldn't just say yes.

And developers are getting the tools to build further: Apple has given developers early access to connect apps to Siri AI, and regular users will have to wait a few months.

What can you do?

Check where the numbers come from. Many of this week's figures come from the vendors themselves, and several features are only in beta or preview. Try agents on small, well-defined tasks where a mistake doesn't cost much, and set clear rules yourself for what they may do without your approval. Keep an eye on whether the new features reach Denmark and what they cost once the testing period ends.