Эпоха болтовни закончена: Claude и OpenAI заставляют нас стать менеджерами

Remember how a couple of years ago we watched mesmerized as the cursor typed out answers to our questions. Back then, the very ability of a machine to sustain a coherent dialogue seemed magical. But the honeymoon with chatbots is over. Developers at Anthropic and OpenAI have realized that the corporate sector is tired of "smart parrots" that can only reason. The latest news around Claude 4.6 and OpenAI's enigmatic Frontier project confirms: we're entering an era of agentic AI, where your role shifts from conversation partner to operations director of a digital team.

The essence of the change is simple, but fundamental. Previously, AI was an oracle: you asked, it answered. Now AI becomes an executor. New models are designed not to produce beautiful paragraphs of text, but to interact with the outside world. They log into your email, open Jira, analyze spreadsheets and independently decide on next steps. This is no longer just "next token generation," it's planning a sequence of actions to achieve a specific business goal. Companies are no longer selling us "smart search," they're selling us outsourcing of middle management.

Why is this happening now? The reason lies in an efficiency ceiling. We've reached a point where simply increasing model parameters doesn't yield explosive growth in answer quality. To justify billion-dollar investments, AI must start delivering direct economic benefit—that is, replacing man-hours in real work processes. OpenAI Frontier and the updated Claude are betting on autonomy. This is a logical step after everyone has played around with image generation and essay writing. Business needs bills to pay themselves and reports to compile without human involvement.

However, the transition to managing agents carries a hidden trap. When you chat with a bot, the cost of its hallucination—an incorrect fact in text that's easy to spot. When you manage an agent, the cost of its error is deleted data, emails mistakenly sent to clients, or a squandered ad budget. We become managers responsible for the actions of our "interns." This requires an entirely different skill set. Prompt engineering, which received so much hype last year, dies before it's born. In its place comes the ability to set clear technical specifications and build a control system.

It's interesting how quickly the workplace landscape will change. If previously copywriters and translators were under fire, now it's time to worry administrators, project coordinators, and everyone whose job consists of moving data from one window to another. Anthropic is already extensively testing the "Computer Use" feature, which lets the model literally move the cursor around the screen. It looks a bit eerie, but such is the new industry standard. We're no longer teaching AI to talk like humans; we're teaching it to work like humans.

In the coming months, we'll see a race for "autonomy." The winner won't be the one whose model writes the best poetry, but the one whose agent makes fewer mistakes when executing routine operations in a browser. This is boring from a hype perspective, but critically important for the real economy. We're left to get used to our role as overseers of a swarm of digital assistants that can be damn effective if you watch over them.

The bottom line: Are you ready to take responsibility for the mistakes your AI agent will make while you're having coffee?