For the past two years, the dominant paradigm of AI interaction has been the chatbot: you type a message, the AI responds. Useful. Impressive. Fundamentally reactive.
That paradigm is changing. A new class of AI systems, called agents, can now take sequences of actions over extended periods of time — browsing websites, writing and executing code, sending emails, booking appointments, analysing documents — with minimal human oversight.
Several companies moved aggressively into agentic AI in 2024 and 2025. Anthropic released Claude’s computer use capability, allowing the model to interact with desktop interfaces as a human would. OpenAI launched Operator, a web-browsing agent integrated into ChatGPT. Startups including Devin (an AI software engineer) built entire products around agentic workflows.
“The analogy I use is that we went from calculators to spreadsheets,” said Harrison Chase, the founder of LangChain. “A calculator does one operation. A spreadsheet does thousands, automatically, in sequence. Agents are the spreadsheet moment for language models.”
The risks that come with autonomous AI action are different in kind from those associated with chatbots. An AI that gives you wrong information is easy to catch and correct. An AI that takes wrong actions — sends an email you did not intend, makes a purchase you did not authorise — creates real-world consequences that may be difficult or impossible to reverse.
Early deployments have produced a mix of remarkable successes and cautionary failures. The companies building these systems are acutely aware of the safety challenges, but the fundamental question of how much autonomy to grant AI systems is one the industry is still working out in real time.
Leave a Reply