Preparing for AI Agents

When the “Parrot” Began to Reason

Andrzej Dragan’s book Quo V-AI-dis highlights a central tension: researchers understand the algorithms used to train neural networks, yet still do not fully understand how those networks later reason.

Models now interpret people and the world in ways that go beyond simple statistical recall. Evidence from science and business in 2025 points to systems entering areas once associated with human intellectual work: reasoning, planning and discovery.

Evidence From Drug Discovery

The Polish source cites three developments from the preceding 12 months. In June 2025, Rentosertib—a pulmonary fibrosis drug for which AI identified the biological target and designed the molecule—completed a successful phase-two clinical trial. In October, another model designed a targeted antibiotic against E. coli. In December, the FDA accepted an AI-generated analysis as scientific evidence.

These examples illustrate the broader shift from fast, intuitive “System 1” answers toward slower, more deliberate “System 2” reasoning.

Polish researchers and engineers have contributed to that shift, including Wojciech Zaremba of OpenAI, Łukasz Kaiser—one of the authors of the Transformer architecture—and Andrzej Dragan. Their framing moves the model from an encyclopedia toward a researcher.

From Chatbots to Agents

A chatbot answers “How do I make a backup?” An AI agent, when properly configured and authorized, can log in, check available resources, create the backup, verify it and prepare a report.

The difference is action. Technology is moving from systems that explain tasks toward agents that execute bounded steps through APIs. The source cites a16z’s Big Ideas 2026 as a signal that infrastructure will increasingly be designed for agents rather than only for people clicking graphical interfaces.

Business use requires explicit permissions, monitoring and approval gates. AI can prepare and execute routine steps within those limits; people approve high-impact decisions and remain accountable for outcomes.

Practical Implications

If models can reason over complex scientific information, they can also support business work such as:

• analyzing documents in a broad context;
• checking business data for inconsistencies;
• drafting and optimizing process designs.

Understanding Instead of Extraction

An agent can analyze a document as a whole, including agreements, acceptance protocols and business dependencies. Multimodal models can interpret layout and context rather than extracting isolated fragments only. A person verifies the interpretation before it drives a decision.

Actions Instead of Clicks

Configured agents can execute steps and exchange data with other tools. They should operate within defined access rights, auditable rules and approval thresholds rather than making uncontrolled decisions.

Creation Instead of Manual Configuration

Natural-language descriptions can produce draft process diagrams, forms and business logic. This lowers the barrier to process design, while process owners review, test and approve each result.

Prevention Instead of Firefighting

Agents can monitor system health, combine signals and identify emerging issues before they become outages. That supports a move from reactive maintenance toward proactive diagnostics, with operators deciding and approving corrective action.

AI as a Controlled Executor

In 2026, AI is moving from an experiment toward performing real work. Organizations can use that opportunity when they replace rigid, isolated forms with governed workflows that define what an agent may do, what must be reviewed and who approves each decision.

The central readiness question is no longer whether AI can generate a plausible answer. It is whether the organization has prepared its data, processes, permissions and human controls for AI-assisted action.

Explore AI-Driven Workflow