MIT Technology Review: Are We Ready to Give AI Agents Real Authority?

The editors of MIT Technology Review have published a closed eBook with a provocative question: are we ready to hand AI agents real power over tasks, data, business processes and, ultimately, the consequences of their actions? The collection brings together the views of leading experts and became one of the publication's most-cited pieces this year.

Agents no longer wait for commands

AI agents are no longer a concept from lab reports. OpenAI, Anthropic, Google, Microsoft and dozens of startups have released systems that use a browser on their own, write and run code, manage files, send emails and interact with external services, all without explicit confirmation of each step.

The agent no longer waits for a command. It receives a goal and moves toward it, making intermediate decisions itself. This shift changes the very nature of the relationship between people and software: we stop being interface operators and become people who delegate responsibility.

There are already examples in production: OpenAI's Operator books restaurants and fills in forms, Claude Computer Use controls desktop applications, and Google Gemini autonomously runs chains of tasks in Workspace. This is not demo mode: agents are running real processes in real organizations.

What the experts say

This is exactly the topic the MIT Technology Review eBook explores, with contributions from leading specialists in AI security, ethics and development. One of the key theses the publication cites:

"If we continue down the current path, we are essentially playing

Russian roulette with humanity."

This is not a line from a dystopian novel. It is the position of practicing researchers who work with agentic systems every day. They single out several specific risk categories:

• Loss of traceability: in multi-agent scenarios, it is hard to determine who made which intermediate decision
• Opacity: when delegating complex tasks, people often do not understand why the agent chose a particular path
• High-stakes errors: financial transactions, medical recommendations, legal documents
• Prompt injection: agents that work with external data can be manipulated through that very data
• Lack of standards: unified industry security requirements for agentic architectures do not yet exist

Between speed and caution

Companies respond to these challenges in different ways. Some, above all Anthropic, bet on multi-layered control mechanisms, constitutional constraints for models and the human-in-the-loop principle for critical decisions. Other players sacrifice caution for speed to market: competitors will not wait, and there has not yet been a major public failure, which is itself a dangerous situation.

Regulators in the US and the European Union are only starting to form a position on agentic systems. The European AI Act partly covers high-risk uses, but it does not yet explicitly regulate agentic architectures as a separate class. While lawmakers refine terminology, the products already operate in real companies and run real processes, often without a clear accountability policy.

What this means

We are entering a phase in which AI agents gain real authority before society has worked out the rules of the game and before businesses have set standards for safe deployment. The questions MIT Technology Review raises in the eBook are not an academic debate. They are an agenda that product teams, legal departments and senior executives making automation decisions already face today.

And the answer to the question "are we ready?" depends largely on how seriously the industry takes the warnings of its own experts.